Predicting Blockchain Energy Consumption: A Step towards a

Sustainable Future
Aryan Rudr Pandit, Arshdeep Singh, Keshav, Kanika
Department of Computer Science Department (Artificial Intelligence)
Chitkara University, Punjab
Corresponding author’s email id: [email protected]

ABSTRACT
The enormous growth of blockchain in this cryptocurrency ecosystem has drawn attention to the
ever increasing power consumption associated with blockchain aggregator systems. It is
estimated that if the same trend continues, the energy consumption will be greater than some
countries' average annual power consumption. This exacerbates environmental issues such as
pollution. In this chapter, we analyze Bitcoins and other cryptocurrencies for energy
consumption from 2009 to 2024. We also attempt to understand the future energy consumption
scenario too, the energy consumed for mining crypto, if we continue to mine Bitcoins (or any
other crypto) in the same way. With the help of various machine learning methods like ARIMA,
XGBoost, and Random Forest, we estimate the parameters like mining difficulty, mining
hardware and also the markets that influence the future energy demands.

It is observed that in order to mitigate energy usage for blockchain, usage of energy efficient
consensus algorithms such as Proof of Stake (PoS), Practical Byzantine Fault Tolerance (PBFT),
Delegated Proof of Stake (DPoS) and usage of clean energy in Bitcoin mining are advocated.
The machine learning capabilities of the blockchain package can further optimize the energy
usage by predicting the demand and distributing the load on the mining site efficiently. The
present advancements in technology and algorithms make it easier to predict the cost for the
process of mining, avoiding wastage of energy.

1. INTRODUCTION
The rapid growth of popularity in the field of blockchain, particularly in the realm of
cryptocurrency mining, has triggered significant concerns regarding the energy consumption as it
takes a whole lot of energy and can have a carbon footprint greater than that of many countries
[Bajra et al., 2024]. The mechanism referred to as Proof-of-Work (PoW) is still the most
commonly used algorithm in numerous block chains even though it is currently known to be
highly consumptive of resources. Large scale mining in the world has seen mining operations use
energy as is being used in several countries. For instance, bitcoin mining itself is expected to take
more than 100 TWh per year by 2024, which will be used to power more than the consumption
of several countries [Kohli et al., 2023].
Such an energy consumption pattern entails disastrous environmental consequences. The
continuous operation of blockchain nodes, driven largely by fossil fuel-based energy like coal
based electrical energy, deepens carbon emissions and accelerates climate change [Qiu et al.,
2023]. In parallel, the current energy networks around the world are experiencing revolutionary
changes with changes in the share of RES (renewable energy sources) including wind and solar
energy. However, integrating blockchain's energy-intensive operations into these more
sustainable systems presents new challenges that are important to be addressed.

In this chapter, we analyze the energy consumption patterns of Bitcoin and other
cryptocurrencies which are viable and have effect on the energy consumption demography right
from their inception in 2009 through 2024 that is shown in figure 2, to identify and predict the
future of energy consumption patterns of blockchain mining.

Consequently, by applying machine learning methods, we predict the future values of energy
consumption with regard to mining difficulty, evolution of mining hardware, development of the
current network system and associated market conditions [Ahmad et al., 2022]. Our analysis will
not only project future energy demands but also explore the results of different strategies for
mitigating this consumption through more energy efficient algorithms, such as Proof-of-Stake
(PoS), Delegated Proof-of-Stake (DPoS), and Practical Byzantine Fault Tolerance (PBFT), along
with the adoption of clean energy sources in mining operations other than fossil fuel based
electric energy. To analyze how these approaches have affected the course of the energy
consumption chart of bitcoin mining [Bada et al., 2021].

By highlighting potential solutions in a form of different mining methods, this chapter aims to
contribute to the ongoing discourse on achieving a more sustainable blockchain ecosystem to
help our tomorrow. in line with global energy transition goals and sustainable development
efforts which can be stable even for the future. In addition, they endeavor to understand
decentralized digital systems of blockchain on energy management such as smart grid
technology and peer-to-peer trading that has the potential to revolutionize energy systems by
increasing reliability, effectiveness and reduction in cost [Wongthongtham et al., 2021]. Hence, it
will be a great foot forward to the suitable future. This problem of high energy consumption due
to crypto mining will not be any harm to our beautiful tomorrow.

The chapter is organised in five more sections. Section 2 explores the work done so far in the
field of blockchain energy consumption. In section 3, we analyze the energy consumption over
the years to depict the bitcoin mining trend in the past decade. Here, we also expound the process
and machine learning models used for predicting the energy requirements. In the next section, we
elaborate the result of predictions using the five Machine learning models. In section 5, we
discuss the implications of the analysis and outcomes of the prediction. In the end, we conclude
with some future possibilities.
2. PREVIOUS WORK

Through exploring various dimensions in this review, the authors seek to look at the positive and
negative sides of blockchain in the energy management and renewable systems domain. In this
section, we cover and explore the concurrent studies that deal with the applications, challenges,
and future directions of blockchain in the energy field.

While blockchain proposes abundant benefits, several studies discuss its challenges in the energy
perspective. Wang et al. (2022) analyse loopholes in present-generation energy systems,
examining technical obstacles like energy efficiency and scalability. They suggest new ideas like
shifting the ledger transaction to Directed Acyclic Graphs (DAGs). Unlike legacy blockchain, in
DAGs based ledgers transactions are organized as a graph instead in sequential blocks. It will
each translate the transaction as a reference to previous blocks and eliminate blocks and miners
as a way to increase scalability and improve its efficiency and apply post-quantum to boost the
character of security. Post Quantum cryptography enhances the security of blockchain by filling
the gap that quantum computers leave in Rivest-Shamir-Adleman (RSA) and Elliptic Curve
Cryptography (ECC) based cryptography. Quantum resistant Algorithms include Lattice based,
hash based and code based Algorithms which does not give any advantage to quantum attacks
and helps for safe integrity and confidentiality of blockchain transactions and data. These
algorithms make blockchain systems future proof for creating secure smart contracts, transaction
validations, and protecting long term sensitive data.

While these advancements can help link blockchain to the rising energy sector’s needs,
regulatory and technical bottlenecks stay. This is because blockchain is not easily compatible
with the actual energy system. Following a similar line of thought, Sahebi et al. (2022) also focus
on barriers blocking the mass adoption of blockchain in the renewable energy supply chain. This
paper addresses key barriers to blockchain adoption in renewable energy supply chains (RESC),
defined as regulatory, institutional, and technological challenges. Compliance and governance
are also made difficult by the lack of adapted frameworks for decentralized systems; the
regulatory hurdles come from that very point. Low industrialization, decentralized knowledge
sharing & lack of managerial expertise along with it create barriers too for institutions to
integrate the blockchain which means a policy change must happen and support system is there
otherwise unexplored benefits do not come from using this technology in renewable energy
administration. The analysis finds that a holistic solution must also include technical solutions, as
examples in the photos showcase, and regulatory solutions.

Despite these problems, few researchers have tried different approaches to investigate solutions
for combat through better management of energy consumption by blockchain technology. An
example of this is a combination of AI and blockchain systems. In 2020, Kumari et al. in their
study say that Blockchain can be combined with AI for efficient energy management.By using
AI-driven classification techniques, the model proposed in the study improves energy demand
prediction and facilitates secure energy trading in smart grids. The integration of AI with
blockchain is proposed as a pathway to create reliable, efficient, and autonomous energy
ecosystems. Luo and Mahdjoubi (2024) extend this integration concept by proposing a
framework combining blockchain with machine learning to manage energy in domestic
buildings. Energy use is optimized and CO2 emissions are reduced via their model, which
consists of LSTM for energy consumption prediction and Particle swarm Optimisation (PSO)
model for scheduling. While we are unable to determine whether the combination of AI and
blockchain would deliver better results compared to other future forecasting methods for energy
management, Jin et al. (2024) has illustrated the potential of such an application combining AI
and blockchain for managing the complex multi energy demands in a decentralized manner
achieving cost reduction and sustainability. With the combination of federated learning and fuzzy
Q learning with blockchain, they provide an optimized model for energy use in hybrid vehicles
to contribute to energy recovery and reduce fuel consumption. In this paper, we illustrate how
blockchain can be adapted to transport applications, especially to the distributed energy resources
such as electric vehicles [Jin et al. (2024)].

Another important aspect of energy management is for decentralized energy communities.

Nykyri et al. (2022) propose a blockchain based solution of balancing energy consumption in
residential areas. According to their study, shared photovoltaic (PV) systems and blockchain
allow for self-sufficiency and efficient energy management. Self-consumption rates are improved
and energy surpluses are redistributed amongst community residents not back to the utility grid.
It is a means to show how blockchain can enable local energy trading in order to improve and
optimize the energy sharing in multi residential buildings. Gill et al. (2019) believes blockchain
can provide cloud computing service which supports integration with IoT and AI within
decentralized applications. Secondly, their review of 140 studies highlight how blockchain, IoT,
and AI can solve the security and efficiency issues for cloud based energy applications.
Proposing the framework, the authors emphasize blockchain’s capacity to ensure data flows
security and ensure energy resilience of cloud computing, key to decentralized systems, and
Andoni et al. (2019) consider wider impact of blockchain in the energy market. Their review of
blockchain projects highlights technology’s promise to promote transparency, secure
transactions, energy democratization especially for the support of small renewable generators.
This sheds light on the regulatory and scalability hurdles for block chain to transition to mass
adoption in the energy space [Andoni et al. (2019)].

In recent works, Ma et al. (2024) addresses how blockchain technology is playing an important
role in the energy sector, analyzing historical trends in energy consumption. In the work, they
demonstrate the patterns in the blockchain, which can broadly point to the direction and
dominant trends. In addition, the study reveals many trends in the integration of blockchain in
the energy sector. Approaches to research have shifted from engineering to a more management
oriented bent, with an increasing focus on governance and market structure. Among them are
blockchain’s support of decentralized energy transaction systems, renewable energy adoption and
smart grid functionality. But there continue to be challenges, business risks, regulatory delays,
and high energy consumption. Fragmented research themes are highlighted in order to initiate
more macro level/holistic studies to address these issues in order to enable future research in the
energy blockchain space. Zhao et al. (2024) on the other side explores blockchain’s role in
localized energy trading, especially within smart communities. This study unveils a
blockchain-enabled framework for peer-to-peer energy exchanges in prominent residential areas.
Results determine that these decentralized models enhance self-sufficiency, allow surplus energy
to be sold, and reduce dependency on centralized grid upgrades. This research and findings
highlight blockchain's capability to create sustainable, autonomous energy communities,
supporting clean energy distribution and promoting energy democratization.

While these studies help illuminate blockchain applications to energy, some parts remain
underexplored. For instance, Zhao et al. (2024) advocate on further investigation of economic
feasibility for large scale decentralized energy markets and lifecycle costs of renewable
infrastructure alike. Likewise, Kumari et al. (2020) suggest enhanced cybersecurity protocols to
deal with privacy related issues in the AI integrated blockchain systems. According to Wang et
al. (2022), regulatory frameworks that encompass blockchain’s decentralization are required in
global energy markets, and Luo and Mahdjoubi (2024) recommend testing blockchain and AI
integration in different energy settings to examine the dynamics of the interplay between these
technologies. Secondly, future research can also consider computational costs and infrastructure
scalability to keep blockchain compatible with the quickly changing demands of the energy
sector.

3. ANALYSIS AND PREDICTION OF ENERGY CONSUMPTION

The primary objective of this research is to predict Bitcoin mining energy consumption
accurately. By utilizing a range of machine learning models, this study aims to minimize energy
usage forecasts and identify future consumption patterns that can guide sustainable practices in
blockchain mining.

The study used two datasets on energy consumption related to Bitcoin mining. The first dataset
(4159 x 16) (Owusu et al., 2022) recorded daily energy metrics, including estimated energy
usage and various energy footprints. To improve the accuracy and efficiency, a second dataset
(138 x 31) (Owusu et al., 2022) was created by aggregating this data monthly. The monthly data
provides smoothed and more manageable insights for model training and testing.

From the plethora of machine learning models, we choose 3 most efficient models to predict the
upcoming range of energy consumption due to blockchain mining- ARIMA (Autoregressive
Integrated Moving Average) whose result can be seen in our table 1, XGBoost (Extreme
Gradient Boosting) on table 2 and Random Forest on table 3. ARIMA is a statistical model
known for its unique capability in time series forecasting, especially when handling stationary
data [Zhang, 2003]. Similarly, XGBoost is also good at handling complex relationships in data
and for its applicability in structured, temporal datasets [Shi et al., 2022 ]. The third model we
apply is Random Forest. Random forest helps avoid overfitting and performs well on
high-dimensional datasets [Ghosh and Cabrera, 2021]. The models are evaluated using the
following three error metrics: Mean Absolute Error (MAE), which measures the average error
magnitude, indicating general accuracy. The formula is given by eqn(1)

MAE = 1/n {Σ(i=1)^n |Actual(i) − Forecast(i)|}, …eqn(1)

where "n" is the total number of data points in the dataset, Actual(i) represents the actual
observed value at the ith data point, "Forecast(i)" is the predicted or forecasted value at the ith
data point, Σ (Summation) represents the sum of the absolute differences between actual and
predicted values across all data points from 1 to n [ArunKumar et al., 2021; Khashei and Bijari,
2010].

Similarly, Root Mean Squared Error (RMSE) highlights larger errors, giving insight into
significant prediction deviations. RMSE is calculated using eqn (2)

RMSE = sqrt{1/n Σ(i=1)^n (Actual(i) − Forecast(i))^2} …eqn(2)

where "n" is the total number of data points, "Actual(i)" is the observed value, "Forecast(i)" is
the forecasted value, and Σ represents the sum of the squared differences between actual and
predicted values across all data points [Wen et al., 2022].

Another metrics is -Mean Absolute Percentage Error (MAPE) normalizes errors as a percentage,
allowing relative accuracy measurement, and is given by eqn(3)

MAPE = 1/n {Σ(i=1)^n |(Actual(i) − Forecast(i))| × 100 / Actual(i)} ...eqn(3)

where "n" represents the number of data points, "Actual(i)" the observed value, "Forecast(i)"
the forecasted value, Σ the summation across all data points, ∣ ∣ indicates absolute error, and
×100 converts the error to a percentage [Dash et al., 1998].

3.1 Model Selection

3.1.1 ARIMA Model

ARIMA(Autoregressive Integrated Moving Average) refers to an autoregressive, integrated and

moving average which is one of the most common methods of forecasting the general value of
observations in time. It is defined by three key parameters: These are (p), (d), and (q) estimated
by using the grid searching method as well as by the ADF stationarity test. The parameter p is
the number of lagged observations or autoregressive terms in the model and the notation d is the
amount of differencing applied to the dataset in order to make it stationary. The value of q, in the
context of moving average indicates the number of past forecast errors incorporated in the
model. In practice, p and q are often integers within the region of 0 to 5 as increasing these
values beyond this precludes a lift in model performance due to over-fitting as a result of
increased model complexity. The value of the differencing parameter d is often chosen in interval
[0, 2], more frequently differentiated data can smooth out some useful features of time series.
Thus, by choosing the appropriate values of the given parameters within these intervals, there is
possibility to analyze the character of the time series without the introduction of the large
additional complexity and without the loss of important trends.

3.1.2 XGBoost Model

XGBoost is a gradient boosting in which the models are built for accuracy on non-linear,
structured data related or connected data sets. Parameters: Hyperparameters include; number of
trees, number of iterations, learning rate, max depth of the tree = (number_of_trees =
1000),(rate_of_learning = 0.01), (maximum depth of the tree = 5). In XGBoost, these parameters
play important roles in controlling the training process and performance of the model: The basis
of the ensemble model is the number of trees and corresponds to the estimators, n_estimators =
1000.The larger value implies here that the model has the ability to capture a feature of a higher
order and however, the time taken in training is more and always prone to
overfitting..parameters: (number_of_trees = 1000),(rate_of_learning = 0.01), (maximum depth of
the tree = 5). In XGBoost, these parameters play important roles in controlling the training
process and performance of the model: The number of trees that form the ensemble model is the
estimators and is defined by n_estimators = 1000. The larger value means that the model has the
ability to learn higher order features and however, the time taken for training is more and may
lead to overfitting. By setting the parameter n_estimators like 1000 the model can create an
instance of up to one thousand trees which will utilize mistakes of previous trees in improving
performance. In the XGBoost, the default value of the learning_rate = 0.01 which bears the
shorthand ‘eta’, refers to the weights allocated to individual trees in a model. They fixed a
standard value of 0.01 for each so that each step is small and the learning rate is therefore low
but is able to generalize. But, when learning rate is low, the number of trees to give a good result
are many, default is 100, n_estimators = X The maximum tree depth an tree is set to have is at
max_depth = 5; this usually ranges from 5 to 30. Deeper model enables modeling higher level
interactions but at the same time incurs overfitting. With the help of an example when
max_depth=5, the main idea is explained that it helps to set limitations for the depth of the tree
and, therefore, provides a sufficient amount of depth in between considering every incoming
feature separately and stressing on this particular feature only excessively.
3.1.3 Random Forest Model

Random Forest enhances the applicability of decision trees in that it allows an ensemble of
decision trees to make a prediction in order to increase accuracy but reduce bias. Parameters: It is
also possible to adjust the settings most important parameters: (n_estimators = 100) and
(max_depth = 10). In a Random Forest model, these parameters play an important role in
shaping the complexity and robustness of the ensemble: n_estimators = 100 stands for the
number of decision trees needed to establish on the mentioned forest. By setting n_estimators to
100, what it means is that the proposed model is 100 trees which are developed from a randomly
chosen data sample as well as features. If more trees are created, accuracy increases and there’s
less of an opportunity for a single tree to over-fit on the data, but computation time will be
higher. The Random Forest with 100 trees as trained is sufficient of getting the optimal balance
between the accuracy of the training and time being taken, while max_depth = 10 gives the
number of levels that the trees can grow to, up to 10 levels, so that each tree is able to capture
moderate complexity in the patterns. This limited depth helps control the formation of deep trees
which could map noise in the data. A depth of 10 is used in the process of effective pattern
recognition while overcoming the issues with excessive complexity and overfitting.

4. RESULTS

4.1 ARIMA Model with Dataset 1

The best_order = (5, 0, 5) was the optimal parameters for the ARIMA model with our dataset 1,
this means that the 5 lag observations are used to predict the future values (visibility range), and
there is no need to make difference to get a stationary result (d=0), the last 5 lagged forecast
errors are incorporated in the moving average component (q=5). The computed Mean Absolute
Error (MAE) for the model is 33,225,120,048.92, which indicates that the average of the
differences between the forecasts and the actual values is about 33.2 billion units of energy
consumption, which is equal to the absolute value of prediction errors. A high level of the RMSE
is denoted from 42,120,050,718.79, which depicts that there is a relatively large number of large
absolute residuals in the model since RMSE gives a lot of consideration to the large errors.
Lastly, Mean Absolute Percentage Error (MAPE) is calculated to be approximately 40.99%
which reveals that the average forecast error is voluming approximately 41% of the actual value
which gives the relative measure of the accuracy of the observed data as envisaged by the model
that can be seen in the figure 1 and figure 4.

Although, it can be deduced that the model makes some points about the data it analyzed, the
amounts of forecast errors mean that there is still room for improvement; it is possible to refine
the model, to analyze more features or try different models.
4.1.2 ARIMA Model with Dataset 2

The ARIMA model for dataset 2, optimized with parameters (p=2, d=2, q=3), produced the
following results: Mean Absolute Error (MAE) of approximately $8,610,692,565,418, Root
Mean Squared Error (RMSE) around $10,643,489,877,368 and Mean Absolute Percentage Error
(MAPE) of about 12.81%. Error estimators also show that the result of dataset 2 with parameters
(2, 2, 3) outperforms the results of dataset 1 with a nearly 74% lower MAE and RMSE. These
small movements imply that dataset 1 has much fewer catastrophic errors and is more likely to
predict actual values with less difference in ascertaining data patterns. Also, on relative grounds,
dataset 2 has a much smaller MAPE of 12.81 % in contrast to dataset 1’s MAPE of 40.99 %.
Because an MAPE of below 15% is often viewed as acceptable in energy consumption
forecasting, Dataset 2 is far better suited for this purpose. Lastly, dataset 2 used differencing with
a value of d=2, this was important in making the data stationary before model fitting, something
very important while using ARIMA on data such as energy consumption data. In the same vein,
Dataset 1 that was forced to have d=0 might have been a problem of non-stationarity which
Annual data may not be conveying well and therefore the lower accuracy it was giving in the
forecast.

Dataset 2 with parameters (2, 2, 3) is clearly more efficient as well as accurate for your dataset,
as evidenced by its lower error metrics and improved stationarity handling. This model is better
suited for reliable energy consumption.

TABLE 1: ARIMA RESULTS TABLE

Selected Model ARIMA

Dataset Dataset 1 Dataset 2
Training Data Split 70% 80%
Testing Data Split 30% 20%
ARIMA Parameters (p, d, q) (5, 0, 5) (2,2,3)
Mean Absolute Error (MAE) 33,225,120,048.92 8,613,527,409.47
Root Mean Squared Error
42,120,050,718.79 10,641,729,835.38
(RMSE)
Mean Absolute Percentage
40.99% 12.81%
Error (MAPE)
Significant error
Forecasting Implication Reduced Errors
observed
Fine-tuning or Enhanced Accuracy and Data
Potential Improvements
alternative models Optimization
4.2 XGBoost Model with Dataset 1

For the given dataset, XGBoost model reached a total test MAE of approximately 19.99 * 10^9
that is on average 19.99 billions the model is off from the true value. The Root Mean Squared
Error (RMSE), is around 28.25 billion; indicating that the present model undergoes considerably
more damage from relatively large errors due to the squared nature of the RMSE. The MAPE is
20.21% therefore the model predictions are on average 20.21 % off from the actual values. These
metrics seem to be reasonable, nonetheless they reveal some potential for further refinement
specific to managing larger deviations. Mae and RMSE indicate the error margin between
predicted and actual energy demand values on average and on their mean squared scale. In this
case both the values are high, it means that the gap between the predicted values and actual data
must be quite large. As we can clearly see, the plot of RMSE is greater than that of MAE,
making RMSE considerably larger, which means there is sensitivity to large errors in the dataset.

It is proposed that a MAPE of ~20.21% implies that on average your model makes mistakes of
about 20% in its predictions from the actual values. Nevertheless, this gives a certain level of
predictive ability; an even lower MAPE, ideally below 10-15%, for energy forecasting. Based on
these outcomes, the model accurately estimates the large-scale consumption trends and might
require tweaking, to increase its effectiveness, for exact prediction. The resulting XGBoost
model was characterized by an average MAPE of 20% which is generally acceptable for most
energy forecasting use-cases, yet, a bit imprecise compared to the high-precision models that
might be required at times. Improving the performance might mean tuning the hyperparameters,
using more features with or without transformations and or other categories of models.

4.2.1 XGBoost model with Dataset 2

Evaluation of the model shows that the Mean Absolute Error (MAE) stood at around 19.09
billion implying that the models predictions were off by an average of 19.09 billion. The RMSE
is about 26.21 billion, making it clear that the prediction errors of a larger magnitude are more
damaging due to the use of squared term for RMSE. The last measure, the Mean Absolute
Percentage Error (MAPE) is slightly bigger than the previous one and it indicates to us that the
average error of the model corresponds to 19.69% of the real values. These values indicate better
accuracy where large errors are less than the previous results obtained.
4.2.2 Comparison
The new XGBoost model shows an improvement, with both lower MAE and RMSE(MAE:
Decreased from ~ 19.99 billion to ~19.09 billion for MSE and from ~ 28.25 billion to ~ 26.21
billion for RMSE). These reductions of MAE and RMSE suggest that the new model is generally
making smaller of average errors and less large deviations. MAPE has also reduced from
~20.21% to ~19.69%. These reductions demonstrate that the new model is relatively more
accurate than the actual values in most cases. The new XGBoost model’s general performance
looks to have gotten slightly better across all 3 evaluation metrics (MAE, RMSE, and MAPE)
implying that tweaks were made to the model; perhaps in hyperparameters or the features, made
the model more accurate. Nevertheless, the differences are not so big nor significant. Although
the new model brings some generalization improvement compared to previous one, there is still a
space for tuning or trying to engineer new features that may bring better results.

TABLE 2: XGBoost Result Table

Selected Model XGBoost

Dataset Dataset1 Dataset 2
19.99 * 10^9 (19.99 19.09 * 10^9
Mean Absolute Error (MAE)
billion) (19.09 billion)

26.21 billion
Root Mean Squared Error
28.25 billion
(RMSE)

Mean Absolute Percentage

20.21% 19.69%
Error (MAPE)
R² (Coefficient of
0.85 0.88
Determination)
Hyperparameter tuning,
Hyperparameter tuning
Potential Improvements cross-validation, and
and feature engineering
ensemble methods
4.3 Random Forest Model

4.3.1 Random Forest Model with Dataset 1

The Random Forest model worked really well for Dataset 1. We had a Mean Absolute Error
(MAE) of roughly 34.78 million so on average the model is way off by approximately 35 million
units. As evident from this relatively small error margin, the model thus effectively predicts. The
MAE was recorded at 96.796 million, however, the Root Mean Squared Error (RMSE) fell to
100.145 million. This means that most predictions are correct, but that there were some bigger
ones with not much distortion across the line, making them more expensive for RMSE to ignore.
In spite of that, the RMSE still points to a sufficient model performance for energy consumption
data which can possess high values. With a Mean Absolute Percentage Error (MAPE) of 3.20%,
the model has predicted its values, on average, to be off by about 3.2%. This is a fine result,
given how often energy consumption values vary here, and has the added benefit of being able to
forecast them.

4.3.2 Random Forest Model with Dataset 2

On the contrary, the performance of the Random Forest model with Dataset 2 was not as good.
Dataset 2 had a MAE of 796.7 million which was a much greater amount than Dataset 1
indicating that the model’s predictions are much less accurate to the actual values. This shows
that the model has a greater MAE because it has more tasks in which to accurately predict energy
consumption at this dataset. Even more so, the RMSE for Dataset 2 was 1.35 billion, much
higher than the RMSE for Dataset 1. The fact that large errors drive the prediction for Dataset 2,
and thereby a less reliable model performance, produces this substantial increase. Dataset 2’s
MAPE, or mean absolute percentage error, was 7.99 %, meaning the model was 8 percent off on
average, which is still okay, but not as good as we saw with Dataset 1.

4.3.3 Comparison with Other Models

Random Forest model performs clearly better than other models (XGBoost, ARIMA, Prophet),
in Dataset 1 and Dataset 2. The MAE observed for Random Forest on Dataset 1 of 34.78 M is
overwhelmingly better than the 19.99 B MAE for XGBoost, thus making Random Forest a much
more reliable model for this dataset. Likewise, the quality of the MAPE for Random Forest of
3.20% far outperforms XGBoost’s 20.21% that is Random forest captures the underlying patterns
in a much more visible way. For Dataset 2, Random Forest outperforms XGBoost, with MAE =
796.7 million vs 19.09 billion for XGBoost. Moreover, Random Forest has a better MAPE than
XGBoost (19.69%) and so its performance is also very good.Random Forest model performs
way better than ARIMA. Random Forest results with MAE of 34.78 million is way too superior
over around 19.99 billion MAE of ARIMA for Dataset 1 to estimate how Random Forest does
better in forecasting energy consumption than ARIMA. Random Forest’s MAPE of just 3.20%
which is also lower than ARIMA 20.21%,indicating better accuracy as shown in figure 4.
Similarly, ARIMA is outperformed by Random Forest on Dataset 2 both in terms of lower error
metrics for both MAE and MAPE. Further, in comparison with Prophet, Random forest exhibits
a striking performance leap. The MAE score of Prophet is 26.59 billion for Dataset 1 but
Random Forest scores only 34.78 million, whilst the MAPE scores 41.47% whereas Prophet
scores 3.20%. The Random Forest performs significantly better, and even for the Dataset 2
Prophet drops behind Random Forest significantly on MAE and MAPE.

Overall however, the Random Forest model performs consistently well for both datasets
especially with Dataset 1, having a low MAE, RMSE, MAPE. Its performance on Dataset 2
wasn’t as cutting edge though (unless I’ve missed a wildcard) but even still it performed better
than the other models tested including a rather popular XGBoost, ARIMA and Prophet. Since the
Random Forest model is able to efficiently represent the inherent trend within the data, the model
renders well for energy consumption prediction and its K-Fold cross validation error is contoured
to outperform other models with respect to accuracy and other error metrics.

TABLE 3: Random Forest Result Table

Selected Model Random Forest

Dataset Dataset1 Dataset 2
Mean Absolute Error (MAE) 34,784,137.45 796,700,000

Root Mean Squared Error 1.35 billion

100.145 million
(RMSE)

Mean Absolute Percentage

3.20% 7.99%
Error (MAPE)
MAE indicates smaller prediction Average deviation of ~796.7
Interpretation errors with a better fit to actual million. A relatively low error
values. margin.
Shows better accuracy with smaller
Shows reasonable accuracy with
Efficiency deviations and more stable
occasional large discrepancies.
predictions

Fine-tuning could lead to marginal Further tuning or hyperparameter

Potential Improvements improvements, but the model is adjustments could improve
already quite effective. accuracy.
4.4 Prophet

Prophet is One open source tool that has been made available to the public by Facebook for the
production of good forecasts from time series data namely the seasonal characteristic and from
data of several seasons in this company. Thanks to it, a consumer can easily incorporate holidays
and other seasonality into the system starting from the present moment. However, while applying
the Prophet Model we arrived at approximately 26,585 million of MAE, approximately 31,033
million of RMSE and 41.47% of MAPE. These metrics show the actual values of these predicted
price values are apprehended and quite distorted and hence are not as effective in this sense as
the other models like ‘Random Forest’ and ‘XGBoost’ which have made fairly low rate of errors.
This indicates that there is an opportunity to decrease the error or that other methods of modeling
should be used to develop sound estimations of energy consumption.

4.5 CatBoost

Energy consumption prediction using the CatBoost model is as follows: MAE: 1.80 billion units,
RMSE: 3.22 billion units, MAPE: 96.19%. The high error rates mean that the model is off by an
average of nearly 96 percentage points when making predictions of the test set values against the
predicted values. This seems remarkably bad when compared with other models which have
been tried before, including Random Forest, which has a MAPE of 7.99%. Amidst the outcomes,
it can be concluded that CatBoost is not helpful in this dataset: the further investigation of the
data quality and losing, feature engineering and, potentially, the models’ specifications is
essential to improving the forecast’s precision.
4.6 Result

All models produced distinct results based on the evaluation metrics. Comparisons among them
reveal the Random Forest model demonstrated the lowest errors across MAE, RMSE, and
MAPE. Detailed results are shown below.

Figure 1: Bitcoin energy consumption over time

Figure 2: Bitcoin mined over time

 Figure 3: Bitcoin mining trends

Figure 4: Comparison between error matrix of all the models on our dataset
5. DISCUSSION

This chapter aims to identify and solve the ever increasingly growing energy consumption in
blockchain mining with an accent on cryptocurrencies like Bitcoin. The increased pace of the
adoption and increase of blockchain technology has heightened concerns over the environmental
and economic impact of the energy expensive Proof-of-work (PoW) consensus mechanism. How
much Bitcoin and other cryptocurrencies actually consume in energy, and how this changes over
time, is the focus of this chapter. The chapter examines historical trends and compares the annual
energy consumption of various countries, and draws comparisons of the potential long term
implications if current mining practices continue. One of the key goals of this research is to use
highly developed machine learning approaches, such as ARIMA, XGBoost, and Random Forest
to predict future energy usage scenarios. The first group of these models takes into account
different parameters like mining difficulty, hardware evolution and trend related market models
that are used to make accurate predictions. The intent of these predictions is to inform the
strategy of reducing energy usage in blockchain mining operations. This study is also advocating
for renewable energy sources integration in the mining operations for cutting down on fossil fuel
use and carbon emissions. The paper also reports how blockchain based energy management
systems such as smart grids and peer to peer trading can be so helpful to better manage the
sustainability and efficiency of energy networks.In the end, the main contribution of this chapter
is to provide guidance to the global debate towards creating sustainable blockchain ecosystems
through providing actionable solutions for achieving energy efficiency. By means of analysis and
recommendations, the study aims to harmonize blockchain technology with diverse
environmental and sustainable development targets so that it remains viable and makes a positive
contribution for the future. Second, in this section, I reflect on the results of these models taken
together, particularly their efforts in predicting the energy consumption from two different
datasets. MAE, RMSE, and MAPE —evaluation metrics were used to evaluate the accuracy of
each model. On Dataset 1, ARIMA (5, 0, 5) had high forecast errors with a MAE of 33.2 billion
and RMSE of 42.12 billion and MAPE of 40.99%. The high error values show that ARIMA had
a hard time capturing the complexity of the energy consumption data in this dataset. Though in
Dataset 2, parameters (2, 2, 3) yielded a MAE of 8.6 billion, RMSE of 10.64 billion, and MAPE
of 12.81% the MAE, RMSE, and MAPE were improved using ARIMA. Nevertheless, the
model's performance was better, although it was not optimal in fact for this task, and in particular
handles non stationary time series data. For Dataset 1, XGBoost takes a MAE of 19.99 billion,
RMSE of 28.25 billion, and a MAPE of 20.21% that outperforms ARIMA and yet still have
noticeable errors in RMSE and MAPE. This makes XGBoost a very good model, but it needs
more work on feature engineering and hyperparameters tuning to extract out the patterns of
energy consumption. Dataset 2 was slightly improved by XGBoost with a MAE = 19.09 billion
RMSE = 26.21 billion and MAPE = 19.69%. As before, these incremental improvements
indicate how XGBoost might further optimize, but still could not outperform Random Forest.
Dataset 1, however, was greatly outperformed by Random Forest compared to both ARIMA and
XGBoost, with MAE and RMSE of 34.78 million and 100.145 million respectively and MAPE
of 3.20%. These results demonstrate the Random Forest’s effectiveness in working with
complex, high dimensional data and in accurate prediction of energy consumption patterns. On
Dataset 2, the MAE of the model was 796.7 million, RMSE of 1.35 billion, and MAPE of 7.99%
(slightly worse than ARIMA and XGBoost in accuracy), this is an improvement over the results
for Dataset 1. On both datasets, the model’s capability to capture energy consumption variations
is demonstrated and its performance.

CONCLUSION

This paper aims to address the challenges of accurately forecasting energy consumption in the
context of Bitcoin mining using advanced machine learning models. For this we worked upon
two distinct datasets, enabling a robust comparative analysis of models such as ARIMA,
XGBoost, Random Forest, Prophet, and CatBoost. Evaluation demonstrated that Random Forest
emerged as the most reliable model, with the lowest MAPE, affirming its suitability for this task.
Our studies explore Bitcoin's energy footprints, the majority are theoretical and lack empirical
validation. This paper aims at fine lining the gap by offering a data-driven, practical approach to
energy consumption prediction. This paper provides actionable insights into the minutes of
Bitcoin's energy demands and lays the foundation for future research, emphasizing the scalability
and real-world applicability of machine learning in addressing cryptocurrency-related energy
challenges. In this study, Random Forest was the best choice for energy consumption prediction
for the two datasets because it had the lower MAE, RMSE, and MAPE over ARIMA and
XGBoost.That is, its superior performance comes from its ability to model complex relationships
and interactions between the data. Whereas the non stationary time series data caused the
ARIMA to perform very badly, in terms of their predictive accuracy. XGBoost had promise, but
couldn’t quite achieve the same precision as Random Forest, requiring a little more refinement to
achieve its level of performance. Random Forest’s robustness towards a large and complex
dataset secures its position as the most trustable model for predicting energy consumption in
blockchain mining. We further tried to improve the ARIMA and XGBoost by both tuning
hyperparameters and feature engineering, but they could not solve this specific task sufficiently.
This raises the possibility of future research to enhance these models making them viable
alternatives. Random Forest is the best choice to solve energy prediction issues in mining with
cryptocurrency due to its accuracy and robustness, and the study shows the necessity of
exploring the future of other applicable models to increase accuracy.
REFERENCES
1. Andoni, M., Robu, V., Flynn, D., Abram, S., Geach, D., Jenkins, D., McCallum, P. and Peacock, A., 2019.
Blockchain technology in the energy sector: A systematic review of challenges and opportunities.
Renewable and sustainable energy reviews, 100, pp.143-174.
2. Gill, S.S., Tuli, S., Xu, M., Singh, I., Singh, K.V., Lindsay, D., Tuli, S., Smirnova, D., Singh, M., Jain, U.
and Pervaiz, H., 2019. Transformative effects of IoT, Blockchain and Artificial Intelligence on cloud
computing: Evolution, vision, trends and open challenges. Internet of Things, 8, p.100118.
3. Jin, W., Li, C. and Zheng, M.Y., 2024. Sustainable energy management in electric vehicle secure
monitoring and blockchain machine learning model. Computers and Electrical Engineering, 115, p.109093.
4. Sahebi, I.G., Mosayebi, A., Masoomi, B. and Marandi, F., 2022. Modeling the enablers for blockchain
technology adoption in the renewable energy supply chain. Technology in Society, 68, p.101871.
5. Nykyri, M., Kärkkäinen, T.J., Levikari, S., Honkapuro, S., Annala, S. and Silventoinen, P., 2022.
Blockchain-based balance settlement ledger for energy communities in open electricity markets. Energy,
253, p.124180.
6. Kumari, A., Gupta, R., Tanwar, S. and Kumar, N., 2020. Blockchain and AI amalgamation for energy cloud
management: Challenges, solutions, and future directions. Journal of Parallel and Distributed Computing,
143, pp.148-166.
7. Wang, T., Hua, H., Wei, Z. and Cao, J., 2022. Challenges of blockchain in new generation energy systems
and future outlooks. International Journal of Electrical Power & Energy Systems, 135, p.107499.
8. Zhao, M., Liao, L., Liang, P., Li, M., Dong, X., Yang, Y., Wang, H. and Zhang, Z., 2024. Energy
Blockchain in Smart Communities: Towards Affordable Clean Energy Supply for the Built Environment.
Energy Engineering, 121(8), pp.2313-2330.
9. Ma, C.Q., Lei, Y.T., Ren, Y.S., Chen, X.Q., Wang, Y.R. and Narayan, S., 2024. Systematic analysis of the
blockchain in the energy sector: Trends, issues, and future directions. Telecommunications Policy, 48(2),
p.102677.
10. Luo, X. and Mahdjoubi, L., 2024. Towards a blockchain and machine learning-based framework for
decentralized energy management. Energy and Buildings, 303, p.113757.
11. Li, J., Liu, X. and Shao, X., 2024. Collaborative carbon emission reduction in power supply and demand
entities based on blockchain technology. International Journal of Electrical Power & Energy Systems, 157,
p.109840.
12. Teufel, B., Sentic, A. and Barmet, M., 2019. Blockchain energy: Blockchain in future energy systems.
Journal of Electronic Science and Technology, 17(4), p.100011.
13. Wang, L., Jiang, S., Shi, Y., Du, X., Xiao, Y., Ma, Y., Yi, X., Zhang, Y. and Li, M., 2023. Blockchain-based
dynamic energy management mode for distributed energy system with high penetration of renewable
energy. International Journal of Electrical Power & Energy Systems, 148, p.108933.
14. Boumaiza, A., 2024. A blockchain-centric P2P trading framework incorporating carbon and energy trades.
Energy Strategy Reviews, 54, p.101466.
15. Shahjahan Alam, K., Daiyan Kaif, A.M.A. and Das, S.K., A Blockchain-Based Optimal Peer-to-Peer
Energy Trading Framework for Decentralized Energy Management with in a Virtual Power Plant: Lab
Scale Studies and Large Scale Proposal. AMA and Das, Sajal Kumar, A Blockchain-Based Optimal
Peer-to-Peer Energy Trading Framework for Decentralized Energy Management within a Virtual Power
Plant: Lab Scale Studies and Large Scale Proposal.
16. Madhwal, Y., Yanovich, Y., Coveri, M. and Marina, N., 2024. Blockchain-Enhanced Hydrogen Fuel
Production and Distribution for Sustainable Energy Management. Blockchain: Research and Applications,
p.100229.
17. Guo, S.M. and Feng, T.T., 2024. Blockchain-based smart trading mechanism for renewable energy power
consumption vouchers and green certificates: Platform design and simulation. Applied Energy, 369,
p.123351.
18. Salehi, M.K. and Rastegar, M., 2024. Blockchain-enabled framework for transactive home energy
management with cloud energy storage under uncertainties. Energy and Buildings, p.114416.
19. Bajra, U.Q., Rogova, E. and Avdiaj, S., 2024. Cryptocurrency blockchain and its carbon footprint:
Anticipating future challenges. Technology in Society, 77, p.102571.
20. Kohli, V., Chakravarty, S., Chamola, V., Sangwan, K.S. and Zeadally, S., 2023. An analysis of energy
consumption and carbon footprints of cryptocurrencies and possible solutions. Digital Communications and
Networks, 9(1), pp.79-89.
21. Qiu, J., Zhao, J., Wen, F., Zhao, J., Gao, C., Zhou, Y., Tao, Y. and Lai, S., 2023. Challenges and pathways
of low-carbon oriented energy transition and power system planning strategy: a review. IEEE Transactions
on Network Science and Engineering.
22. Ahmad, T., Madonski, R., Zhang, D., Huang, C. and Mujeeb, A., 2022. Data-driven probabilistic machine
learning in sustainable smart energy/smart energy systems: Key developments, challenges, and future
research opportunities in the context of smart grid paradigm. Renewable and Sustainable Energy Reviews,
160, p.112128.
23. Bada, A.O., Damianou, A., Angelopoulos, C.M. and Katos, V., 2021, July. Towards a green blockchain: A
review of consensus mechanisms and their energy consumption. In 2021 17th international conference on
distributed computing in sensor systems (DCOSS) (pp. 503-511). IEEE.
24. Wongthongtham, P., Marrable, D., Abu-Salih, B., Liu, X. and Morrison, G., 2021. Blockchain-enabled
Peer-to-Peer energy trading. Computers & Electrical Engineering, 94, p.107299.
25. Zhang, G.P., 2003. Time series forecasting using a hybrid ARIMA and neural network model.
Neurocomputing, 50, pp.159-175.
26. Shi, L., Qian, C. and Guo, F., 2022. Real-time driving risk assessment using deep learning with XGBoost.
Accident Analysis & Prevention, 178, p.106836.
27. Ghosh, D. and Cabrera, J., 2021. Enriched random forest for high dimensional genomic data. IEEE/ACM
transactions on computational biology and bioinformatics, 19(5), pp.2817-2828.
28. Owusu, Phebe Asantewaa; Sarkodie, Samuel Asumadu (2022). Dataset on bitcoin carbon footprint and
energy consumption. figshare. Dataset. https://doi.org/10.6084/m9.figshare.19442933.v1
29. ArunKumar, K.E., Kalaga, D.V., Kumar, C.M.S., Chilkoor, G., Kawaji, M. and Brenza, T.M., 2021.
Forecasting the dynamics of cumulative COVID-19 cases (confirmed, recovered and deaths) for top-16
countries using statistical machine learning models: Auto-Regressive Integrated Moving Average (ARIMA)
and Seasonal Auto-Regressive Integrated Moving Average (SARIMA). Applied soft computing, 103,
p.107161.
30. Khashei, M. and Bijari, M., 2010. An artificial neural network (p, d, q) model for timeseries forecasting.
Expert Systems with applications, 37(1), pp.479-489.
31. Wen, Z., Zhou, R. and Su, H., 2022. MR and stacked GRUs neural network combined model and its
application for deformation prediction of concrete dam. Expert Systems with Applications, 201, p.117272.
32. Dash, P.K., Satpathy, H.P. and Liew, A.C., 1998. A real-time short-term peak and average load forecasting
system using a self-organising fuzzy neural network. Engineering Applications of Artificial Intelligence,
11(2), pp.307-316.