Interpretable Machine Learning Approach to Forecast Sustainable Power Generation

Electricity is vital for the progress of society and human civilization, serving as a key energy source that supports the steady growth of the global economy [1]. Over the past 20 years, the electricity system has faced increasing pressure to reduce CO₂ emissions, meet rising demand, provide reliable and affordable services, and sustain economic growth [2]. Global electricity demand is projected to rise from around 20,100 terawatt-hours (TWh) per year in 2013 to between 30,000 and 37,400 TWh per year by 2040 [3]. Currently, electricity generation accounts for roughly 40% of global CO₂ emissions, making it one of the largest contributors to greenhouse gases (GHG) [2]. To mitigate climate change, transitioning to low-carbon electricity sources like wind, solar, water, geothermal, and biomass is essential [4]. Accurately predicting electricity generation from these renewable sources can guide stakeholders toward sustainability and support climate strategies. Machine Learning (ML) models excel in handling large datasets, improving over time by uncovering complex patterns and extracting features from raw data [5]. Various studies shows that ML models can play a pivotal role in low carbon electricity predictions [6]. Although ML models frequently yield accurate predictions, their opaqueness makes it challenging to comprehend the logic underlying their judgments. This is especially troublesome when these forecasts affect important decision-making procedures. To solve this, eXplainable Artificial Intelligence (XAI) techniques offer post-hoc interpretations that enable us to determine which elements have the greatest impact on the model's predictions [7]. As far as we are aware, no prior research has utilized XAI to interpret the results of ML models in low-carbon electricity contexts. The contribution of this study is:

1. We develop a comprehensive approach to predict the production of low-carbon electricity by analysing and contrasting several ML models.

2. We demonstrate the global and local explainability to our best performed model using SHAP and Shapash.

3. We identify the most significant features that impact the performance of the ML models in prediction using global explanations.

The remaining parts of the paper are arranged as follows: Section 2 contains the relevant literature; Section 3 contains the techniques and materials; Section 4 contains the findings; and Section 5 contains the conclusion and suggestions for further research.

This section provides an overview of previous studies on low-carbon electricity prediction, highlighting current approaches and recent advancements in the field.

There is diverse research worldwide with 13,767 publications on low carbon electricity both qualitatively and quantitatively while United States has along contributed 3074 publications in a spanning of quarter century [2]. To address difficulties and opportunities numerous forecasting techniques have been proposed in recent years [8]. Autoregressive integrated moving average (ARIMA), the grey models and the Linear Regression (LR) models are most often used traditional model for forecasting [9]. The main drawback of such model is they influenced by region that is different models require for different region and their pattern of electricity consumption or production [6]. ML is the popular models of forecasting in the quantitative research of low carbon electricity. Researchers have been interested in ML algorithms because of their improved performance and capacity to find nonlinear correlations. ML algorithms have demonstrated encouraging results in wind power prediction, providing a number of benefits over conventional methods. A new technique that uses ensemble learning to anticipate wind power generation one day ahead of time is described, effectively managing curtailment and turbine damage. For effective hyper-parameter tweaking a hybrid forecasting model combines Wavelet Packet Decomposition (WPD), Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) where Sequential Model-Based Optimization (SMBO) is used for combines [10]. In the area of photovoltaic power forecasting, the two most used machine learning techniques are Support Vector Machine (SVM) and Artificial Neural Networks (ANN) [11]. SVM is used in biomass energy prediction to avoid the challenge of regression and classification [12].

In this section, we describe the tools and techniques used in this study. We first use an overview diagram to show how the process works, and then we go into great detail about each individual part.

This section presents the results based on the outlined contributions, starting with the performance of the ML models, followed by the explainability results, revealing the most influential factors in predicting low-carbon electricity production.

The results explain significant differences for various ML models that are listed in table Table 1. In our study RF emerged as the best performing model among others. RF model provides more effective MSE of 7.741, R² of 99.3%, MAE of 1.443 and MSE of 2.782. Low error rates and high R² make this model strongest choice of forecasting low carbon electricity generation. Though surpassed by RF, GBM and XGB also provide significantly more accurate predictions than other models. Additionally, SVM is the worst performer to predict the low carbon electricity generation with an MSE of 1075.046, R² of 8%, MAE of 26.382 and RMSE of 32.787.

The main goal of this study is to predict low carbon electricity generation and suggest the best model for forecasting. RF perform better among all the model with lowest MSE, RMSE, and MAE additionally highest value of R². Only a few additional models exhibit encouraging outcomes; the best two are XGB and GBM. The smooth integration of zero carbon electricity sources, grid stability, and efficient energy management all depend on their accuracy and dependability. Their exceptional precision lessens reliance on fossil fuels and takes one step ahead for zero carbon goal. Then XAI tools such as SHAP and Shapash based on RF model, reveals important features like Electricity from renewables (TWh), Electricity from fossil fuels (TWh) Renewable energy share in the total final energy consumption (%), and Electricity from nuclear. This research admits its limitations, such as the dataset's possible incompleteness and the analysis's lack of consideration of economic and policy aspects. It is essential to use a more extensive data set, investigate economic aspects, and include real-time data in order to further future study.

The submitted work is a unique contribution to the field, not published elsewhere in any form or language. Results are presented clearly, honestly, and without fabrication, falsification or inappropriate data manipulation (including image-based manipulation). Authors adhere to discipline-specific rules for acquiring, selecting and processing data. 

The submitted work is experimental work performed in the laboratory. No human subject or living organism/tissue is involved in this research. 

No human subject or living organism is involved in this research. So, no consent needed to publish is to be shared.