Reliable renewable energy forecasting for climate change mitigation

Motivation

The essential need to strengthen the stability and dependability of RESs, particularly in the context of growing environmental concerns and emergency scenarios, is the basis behind this study. A revolutionary transition to sustainable energy resources is required in order to address climate change and reduce our dependence on traditional fossil fuels (Kabeyi & Olanrewaju, 2022). With the ability to decrease the effects of climate change and guarantee a sustainable energy future, RESs, provide a promising path forward for clean and green energy (Al-Shetwi, 2022). Robust and adaptableRESs are even more crucial in emergency scenarios, such as emergencies or disruptions to traditional energy networks. Precise forecasting of renewable energy output is essential for daily operations as well as for quickly and efficiently handling unanticipated events (Alkabbani et al., 2021). This study aims to combine green and nature-inspired approaches into renewable energy forecasting, driven by the critical need to develop resilient energy systems that can resist disasters and aid in disaster recovery efforts.

Contribution

By addressing important issues and presenting novel approaches, our study significantly advances the area of renewable energy forecasting. The following are this work’s primary technical contributions:

1. Hybrid architecture for projecting renewable energy: We suggest a unique hybrid design to capture intricate interactions between various RESs, combining ResNext (Pant, Yadav & Gaur, 2020) and GRU (Wang, Liao & Chang, 2018) models. In contrast to other methods that frequently concentrate on estimating distinct energy sources independently, our hybrid model takes into account the localized correlations between many sources. Because of the complex interrelationships between various energy components, this design yields more precise estimates of overall power generation.

2. Implementing PCA for feature extraction: We use PCA for feature extraction in order to increase the dataset’s temporal dimension and interpretability. Effective feature selection is aided by PCA, which decreases the dimensionality of the data while keeping important information. Improved model performance is made possible by this contribution, which makes it possible to express linear time-series features more effectively.

3. Rigorous performance evaluation standards: Using rigorous measures such as R-squared, explained variance, mean absolute error (MAPE), root mean squared error (RMSE), and RMS logarithmic error, we conduct a thorough performance review. This comprehensive examination guarantees a detailed appraisal of the model’s precision and accuracy, enabling a trustworthy comparison with the most recent versions. A more complex understanding of the model’s predicting ability is made possible by the use of several measures.

4. Superior predictive performance for hourly energy generation: Our proposed ResNeXt-GRU-MJA model outshines existing state-of-the-art methods by demonstrating a remarkable 15% improvement in hourly forecasting accuracy. This superior performance is substantiated through normalized measures of RMSE, MAPE, and MAE specifically tailored for hourly predictions. The model’s effectiveness in capturing short-term energy variations positions it as a valuable tool for applications requiring precise hourly forecasting.

Article organization

The article is structured into several sections. ‘Related Work’ provides a systematic review of existing methods developed by researchers. In ‘System Model’, we present the innovative approach proposed to enhance the current work. ‘Simulation and Results’ delves into the experimental simulation, offering insights into the practical implementation of the methodology. The conclusion is then presented to summarize key findings and contributions.

Dataset collection and description

To ensure the sustainability of our data preparation processes, we logically separated the dataset into training and testing sets using an 80/20 ratio. This separation enables us to make the most use of the data by dividing it up and assigning a significant amount to model training and a separate chunk to performance assessment. We aim to maintain the integrity and dependability of the assessment procedure. The dataset utilized in this study is a combine of two datasets from Kaggle (Afroz, 2023) and Anikannal (2023), totaling 8,760 occurrences. With the help of these repositories, researchers can utilize and analyze a wide range of publicly available datasets for a variety of research applications. We actively encourage transparency and reproducible in our research by utilizing datasets from various sources, making it possible for others to validate or expand upon our findings. Table 2 provides a description of the applicable dataset.

Pre-processing

To provide reliable and precise outcomes, data preparation is essential (ShobanaDevi et al., 2021). In order to prepare the data for analysis and modeling, careful handling of outliers and missing values at this crucial stage, removal of unnecessary attributes, and utilization of feature extraction methods like PCA are performed. Since errors and outliers may significantly compromise the accuracy of the analytic results, fixing them is an essential part of the pre-processing stage. To preserve the integrity of the data, strong methods are employed to detect and manage missing values. This covers interpolation techniques like imputation based on regression or mean (Eq. (1)) or median (Eq. (2)). By using these strategies, the aim is to promote sustainable data practices by maintaining the dataset’s completeness and reducing the impact of missing values on subsequent studies.

To handle missing values, two strategies are employed: mean imputation and median imputation. The mean imputation, expressed by Eq. (1) (Josse & Husson, 2012), estimates the missing value ${\hat{x}}_i$ by averaging the available data within the corresponding feature: (1)${\displaystyle {\hat{x}}_i=\frac{\sum _{j=1}^n{x}_j}{n}.}$

Similarly, the median imputation, depicted in Eq. (2) (Josse & Husson, 2012) as Eq. (2), utilizes the median of the available values in the feature: (2)${\displaystyle {\hat{x}}_i=\text{median}\left(x_1,x_2,\dots ,x_n\right).}$

Outlier detection is then performed using the z-score (Eq. (3)) (Josse & Husson, 2012), allowing the identification and appropriate handling of outliers: (3)${\displaystyle z=\frac{x-\mu }{\sigma }.}$

In this case, the data point is denoted by x, the average by µ, and the standard deviation by σ. To enhance the classification model’s efficiency, redundant features are eliminated based on their importance, calculated using XGB (Eq. (4)) (Josse & Husson, 2012): (4)${\displaystyle {\text{importance}}_i=\text{gain}.}$

In this equation, importance_i signifies the importance rating of feature i, and gain indicates the improvement in the model’s performance achieved by utilizing feature i.

Additionally, during preprocessing, irrelevant data is identified and removed, reducing computational overhead and ensuring focus on the most pertinent information for analysis. This meticulous approach addresses missing values and outliers, eliminates redundant features, and streamlines the dataset, contributing to its quality and suitability for subsequent analysis and modeling.

Exploratory data analysis

Our work integrates sustainable practices into the exploratory data analysis (EDA) phase with the goal of accurately anticipating renewable energy to combat climate change. Using visual analysis and mapping approaches, our complete EDA prioritizes sustainability principles and aims to obtain significant insights into the dataset (Milo & Somech, 2020).

The EDA strategy focuses on using mapping methods to identify connections between category data and other parameters. We begin a process of inquiry that improves our comprehension of distribution, ratios, and interactions among categorical data by visually analyzing intricate connections and interrelationships. Visual tools that support sustainable behaviors, such graphical representations and contingency tables, help people make meaningful decisions and findings.

Furthermore, a comprehensive visual examination of the dataset investigates patterns, abnormalities, and relationships. We examine the distribution, diversity, and interactions between variables using a range of visualizations, such as data variations, scatter visualizations, patterns graphs, repetition charts, and statistical box plots. Identifying anomalies or outliers, revealing hidden patterns, and evaluating variables for relationships or dependencies are all made possible by sustainable visualization approaches. The results of our EDA show possible quality problems and offer insightful information about the properties of the dataset. This enhanced understanding creates an effective basis for further research and well-informed decision-making. Our decision-making process is guided by sustainable EDA methodologies, such category mapping and visual analysis, which also help to build a trustworthy classification model for predicting the output of renewable energy (Milo & Somech, 2020).

Proposed renewable energy forecasting model: ResNeXt-GRU-MJA

Our research introduces the ResNeXt-GRU-MJA model, a hybrid architecture tailored for precise renewable energy forecasting, aligning with our commitment to sustainable practices. This model seamlessly integrates the ResNeXt and GRU architectures with the optimization capabilities of the MJA, offering a comprehensive and customized solution for renewable energy prediction. The internal structure of the ResNeXt-GRU-MJA model is illustrated in Fig. 2.

Feature extraction with resNeXt (Li, Zhu & Zhu, 2023): The model’s prediction process begins with raw renewable energy data as input. the ResNeXt component processes the information through convolutional layers, batch normalization, and ReLU activation for each path i within the ResNeXt block.

Path Cardinality and Concatenation: To capture diverse features, data is strategically partitioned into multiple paths based on cardinality. The outputs from these paths are concatenated to form a holistic feature representation. (5)${\displaystyle \text{Concatenation}=\left[ReLU1,ReLU2,ReLU3,ReLU4\right].}$

Sequential modeling with GRU: The concatenated features traverse a GRU layer, introducing a temporal modeling dimension. This empowers the model to decipher sequential dependencies and evolving patterns within the renewable energy data.

Hyperparameter tuning with MJA: The MJA is utilized to fine-tune the model’s hyperparameters, which are crucial for convergence, generalizing, and performance. The model’s performance on a validation dataset is used to inform the exploration and updating of hyperparameter values in this iterative process. The hyperparameters and their assigned values are shown in Table 3.

Hybrid classification model algorithm (ResNeXt-MJA): The proposed hybrid model, ResNeXt-GRU-MJA, incorporates the ResNeXt model with the optimization capabilities of the MJA. This hybrid approach offers a customized solution for renewable energy forecasting, enhancing accuracy and effectiveness.

Algorithm 1  shows the overview of our proposed work. The hybrid classification model involves pre-processing, exploratory data analysis (EDA), optimization using MJA, and performance evaluation. It ensures the model is fine-tuned, taking into account specific characteristics of the renewable energy dataset. The ResNeXt-MJA model represents a tailored and sustainable solution for accurate renewable energy forecasting, contributing to the broader goal of mitigating climate change.

 
_______________________ 
Algorithm 1 Hybrid classification model for renewable energy forecasting_____ 
  1:  procedure RENEWABLEENERGYFORECASTING(Renewable Energy Data) 
  2:       Pre-processing: 
 3:           Handle missing values and outliers using robust techniques 
  4:           Remove redundant features 
  5:           Perform feature selection using ResNeXt 
  6:       Exploratory Data Analysis (EDA): 
 7:           Conduct descriptive analysis 
  8:           Map categorical features in relation 
  9:           Perform graphical analysis 
10:       Optimization Process using MJA: 
11:           Initialize a population of hyperparameters 
12:           Evaluate performance using ResNeXt 
13:           Select the fittest hyperparameters for reproduction 
14:           Apply mutation and crossover operations 
15:           Refine hyperparameters through a growth phase 
16:           Conduct competition for survival 
17:           Repeat for a predetermined number of generations 
18:       Performance Evaluation: 
19:           Split data into training and testing sets 
20:           Train ResNeXt-GRU-MJA model on the training data 
21:           Evaluate performance using various metrics 
22:           Conduct statistical analysis 
23:           Calculate computational complexity measures 
24:       Output: Hybrid Classification Model (ResNeXt-GRU-MJA) for Renew- 
     able Energy Forecasting 
25:  end procedure___________________________________________________________________    

Split cardinality: The proposed ResNeXt-GRU-MJA model introduces split cardinality in its ResNeXt blocks. This involves dividing the cardinality (number of groups) into smaller fractions, allowing the model to capture more diverse and localized correlations among different energy sources. The split cardinality enhances the model’s ability to understand complex relationships within the dataset. The split cardinality in our ResNeXt blocks is mathematically represented as follows: (6)${\displaystyle {\text{Cardinality}}_{\text{split}}=\frac{\text{Total Cardinality}}{\text{Number of Splits}}.}$ This division allows the model to capture diverse correlations among energy sources, enhancing its ability to understand intricate relationships.

Attention mechanism: Our model incorporates an attention mechanism, which focuses on capturing relevant information from different energy sources during the modeling process. The attention mechanism assigns varying levels of importance to different parts of the input data, allowing the model to dynamically adjust its focus based on the significance of each source. This attention mechanism contributes to the model’s ability to effectively capture localized correlations and improve forecasting accuracy. The attention mechanism is modeled as a weighted sum in our GRU cells. Given an input sequence (X′), hidden states (H′), and attention weights (W′), the weighted sum is calculated as: (7)${\displaystyle \text{Weighted Sum}=\sum _{i=1}^{N^{\prime }}{W}_i^{\prime }\cdot H_i^{\prime }.}$ Here, (N′) represents the number of elements in the sequence, and $\left(W_i^{\prime }\right)$ denotes the attention weight assigned to each element. The attention mechanism dynamically adjusts these weights based on the significance of each energy source, contributing to improved forecasting accuracy.

Comprehensive performance evaluation of ResNeXt-GRU-MJA hybrid model

We conducted a comprehensive and rigorous evaluation procedure on our novel ResNeXt-GRU-MJA hybrid model to determine its level of competence in renewable energy forecasting. The assessment process was multiple phases and intended to provide an extensive understanding of the model’s advantages and disadvantages.

A diverse set of evaluation metrics was employed to thoroughly assess the ResNeXt-GRU-MJA model’s performance:

• Root mean squared error (RMSE): Gauging the average magnitude of errors between predicted and actual values, providing a holistic measure of prediction accuracy.

• RMS LOG error: A logarithmic application of the RMSE, beneficial for datasets with a wide range of values.

• Explained variance (Exp Variance): Quantifying the proportion of variance in predicted values, offering insights into the model’s explanatory capacity.

• R-squared (R₂): Representing the predictability of the dependent variable from the independent variables. A higher R-squared indicates a superior fit.

• Mean absolute percentage error (MAPE): Evaluating the percentage difference between predicted and actual values, providing a normalized assessment of prediction accuracy.

• Benchmarking against state-of-the-art algorithms: Apart from assessing the ResNeXt-MJA model separately, we also carried out a comparison study with industry-leading algorithms including ARIMA, CNN, VGG, NB, DenseNet, and decision trees. This thorough benchmarking made it easier to comprehend our model’s performance in the context of the larger renewable energy forecasting field.

• Rigorous statistical analysis: We conducted a thorough statistical study to strengthen the validity of our findings. This ensured the validity of observed differences and the strength of our results. It also included confidence interval estimates and hypothesis testing.

The ensuing sections present detailed results and discussions, shedding light on the ResNeXt-MJA hybrid model’s predictive capabilities and its comparative standing in the realm of renewable energy forecasting methodologies.