SEG forecasting has attracted significant research efforts owing to the rise of PV systems and smart grid installations. The application of various machine learning techniques such as Support Vector Regression (SVR), Random Forest (RF), Artificial Neural Networks (ANN), Gaussian Process Regression (GPR), Decision Tree (DT), and Gradient Boosting for forecasting has shown great success in accurately modeling the nonlinear dependency between meteorological features and power production [22][23][17][4]. Particularly, SVR is still one of the most popular forecasting methods due to its good generalization ability, whereas RF and ANN yield high forecasting accuracy regardless of climate [17].

Machine learning models and optimization algorithms have been combined to increase prediction accuracy. To optimize model hyperparameters and lower forecasting errors, methods like Particle Swarm Optimization (PSO) and a Hybrid Improved Multi-Verse Optimizer (HIMVO) [22], Genetic Algorithms (GA/GASVM) [23], Ant Colony Optimization (ACO) [24], the Whale Optimization Algorithm (WOA) [25], and Bayesian Optimization [21] have been used. Despite the fact that these hybrid approaches increase prediction accuracy, most research ignores the impact of feature selection on forecasting efficiency in favor of parameter tuning.

Deep learning models, comprising Long Short-Term Memory (LSTM) networks [26], Gated Recurrent Units (GRU) [27], Convolutional Neural Networks (CNN) [28], and Transformer architectures [29], have recently demonstrated excellent performance in solar forecasting by effectively learning temporal and nonlinear patterns from large datasets. However, these models' application in computationally constrained situations is limited since they often demand large amounts of training data, higher computational resources, and longer training times [30].

An efficient preprocessing method for increasing forecasting efficiency is feature selection. Data dimensionality is decreased while model generalization is enhanced by techniques including Pearson Correlation (PC) [17], Recursive Feature Elimination (RFE) [31], Random Forest Feature Importance (RFFI) [19], Mutual Information [32], and wrapper-based approaches [20]. However, current research usually assesses a single feature selection method or combines it with a small number of forecasting models, offering little information on the relative efficacy of several feature selection approaches.

SEG forecasting has come a long way, but there are still a number of obstacles. The majority of current research focuses on optimizing models or hyperparameters while paying little attention to systematic feature selection. There are also few comparative analyses of various feature selection methods under the same experimental conditions. Additionally, traditional criteria are frequently used to evaluate predictive ability, while advanced regression measures are rarely taken into account. These constraints drive the creation of computationally effective forecasting frameworks that use small feature subsets to retain high predicted accuracy.

This study suggests a feature-selection-driven forecasting paradigm that compares Random Forest Feature Importance (RFFI), Recursive Feature Elimination (RFE), and Pearson Correlation (PC) in order to address these issues. An optimal Radial Basis Function Support Vector Regression (RBF-SVR) model is created after four machine learning regression models are trained with similar feature subsets. Both traditional and sophisticated regression performance indicators are used to thoroughly assess the suggested architecture. The current work is positioned in relation to representative recent investigations, which are summarized in Table 1.

The experiments were performed based on the photovoltaic data set from the Desert Knowledge Australia Solar Centre (DKASC), which is an open-source data set available for research use. This data set includes actual measurements from Site 28 at the DKASC Alice Springs site, which is a First Solar CdTe (thin-film) solar power plant. Table 2 presents the specifications of Site 28.

The dataset includes 13 predictor variables, a target variable (Active Power), and a timestamp field, recorded at an interval of five minutes over several years, resulting in 1,281,034 total observations after cleaning. However, two out of the 13 predictor variables recorded (Current Phase Average and Active Energy Delivered/Received) have been excluded from the candidate feature set because they represent electrical equivalents of the target variable and thus do not qualify as predictor variables, nor will they be available at the time of forecasting. The ten remaining candidates are listed in Table 3 below.

Reliable forecasting depends on high-quality input data. Three preprocessing steps were performed before model development.

Handling Missing Data: The missing data, which occurred due to disruptions in sensors and/or communication, were filled through median imputation calculated for both training and testing datasets separately so that no information was leaked from the testing dataset to the training one.

Data Filtering: Occasionally, anomalous values in sensor measurements arise from hardware constraints or environmental disruptions. Prior to model training, filtering techniques were used to eliminate noisy and inconsistent readings.

Feature Normalization: Meteorological variables possess different numerical ranges, which may negatively influence machine learning algorithms. Min–Max normalization was applied to scale each feature into the interval [0,1], accelerating model convergence and preventing variables with larger magnitudes from dominating the learning process, as defined in Equation (1):

x_scaled = (x - min(x)) / (max(x) - min(x)) (1)

The dataset is sorted chronologically, and in the case of large-scale model comparisons presented in Section 4, a random sampling of 60,000 rows is used, where a 70:30 split ratio is used between train and test sets (42,000 training and 18,000 testing samples) in order to maintain computability and validity of the results on such large-scale datasets. This method is being used for all of the reported experiments.

The most informative meteorological variables for SEG forecasting were determined through feature selection. The four top-ranked characteristics from the initial set of thirteen were chosen by each of the three complementary methods that were examined, each of which assessed feature significance from a different angle.

After feature selection, each subset of features was assessed using four popular regression algorithms, which include Linear Regression (LR), Random Forest (RF), Support Vector Regression (SVR), and Multilayer Perceptron (MLP). The four baseline models are briefly introduced below; their performances are compared in Section 4, where Random Forest is found to be the best among them. The Optimized RBF-SVR algorithm proposed in the paper is described in detail in Section 3.5.

While traditional SVR has shown competitive performance in solar energy forecasting, feature quality and hyperparameter choice have a significant impact on prediction accuracy. This section combines feature selection, polynomial feature transformation, feature normalization, and randomized hyperparameter optimization into a single forecasting pipeline.

The reason why RBF-SVR was preferred over deep architectures (such as LSTM, GRU, Transformer-based models; see Section 2.2) is two-fold. Firstly, kernel-based models such as SVR usually need considerably fewer training samples to perform well compared to deep architectures. This is due to the fact that this research evaluates models on a small 4-5 feature set for each feature selection algorithm, rather than large multivariate sequences as input. Secondly, training deep architectures usually takes considerably more computational resources and time to be fine-tuned. The training times for the Optimized RBF-SVR presented in Section 4.4 demonstrate that, on average, it needs only 21.25 seconds to complete the training, while in the literature (see Section 2.2) LSTM- and Transformer-based solar forecasting models require much more time for training. The speed of training is explained by the fact that the objective of this research is resource-efficient forecasting, not absolute accuracy regardless of the computational time needed for that.

The proposed forecasting framework was implemented in Python 3.11 using the Google Colab environment, with additional validation performed in a Kaggle Notebook. Data preprocessing, feature selection, machine learning implementation, and performance evaluation were carried out using the Scikit-learn, NumPy, Pandas, and Matplotlib libraries. Hyperparameter optimization for the proposed RBF-SVR model was performed using RandomizedSearchCV on the training data only. Table 7 summarizes the experimental configuration.

Model performance was assessed using eight complementary regression metrics: MSE, RMSE, MAE, MAPE, MedAE, EVS, Huber Loss, and R². Error-based metrics quantify prediction accuracy, whereas EVS and R² evaluate the goodness-of-fit between predicted and observed photovoltaic power output. Their mathematical definitions are given below.

MSE = (1/n) Σ(yᵢ − ŷᵢ)² (7)

(8)

MAE = (1/n) Σ|yᵢ − ŷᵢ| (9)

MAPE = (1/n) Σ|(yᵢ − ŷᵢ)/yᵢ| × 100 (10)

MedAE = median(|y₁ − ŷ₁|, |y₂ − ŷ₂|, …, |yₙ − ŷₙ|) (11)

EVS = 1 − Var(y − ŷ) / Var(y) (12)

R² = 1 − Σ(y − ŷ)² / Σ(y − ȳ)² (13)

Lδ(a) = ½a²  if |a| ≤ δ;   Lδ(a) = δ(|a| − ½δ) (14)

where yᵢ is the actual value, ŷᵢ is the predicted value, n is the number of observations, ȳ is the mean of the observed values, and δ is the Huber threshold parameter. Lower values of MSE, RMSE, MAE, MAPE, MedAE, and Huber Loss indicate better prediction accuracy, whereas higher values of EVS and R² represent superior model performance, as defined in Equations (7) to (14) above. Table 8 summarizes the objective of each metric.

MAPE is not included in the results in Tables 9, 10, and 11 due to its being undefined for yᵢ = 0 and mathematically unstable when yᵢ ≈ 0, which is common for Active Power in the evening and dawn/dusk transitions in the dataset at hand. Even upon restriction of MAPE calculation to daytime samples (Active Power > 0.05), the MAPE in this experiment continued to be in the hundreds of percent, owing to small denominators that still occurred during the sunrise/set ramp. It is a known flaw of MAPE for signals that pass through or approach zero, which is the reason why this paper uses MedAE and Huber Loss (which behave fine regardless of proximity to zero) as its main robust error measures.

Figure 3 contains all seven metrics listed in Table 9, separated into two panels to keep metrics with conflicting meanings on one coordinate axis separate: error metrics (MSE, RMSE, MAE, MedAE, Huber Loss) for which low values are good, and goodness-of-fit metrics (R-squared, EVS) for which high values are good. There is no model surpassing R² = 0.57 for the PC-selected features, and the Optimized RBF-SVR performs about equally well to the Linear Regression using this particular set of features (MSE = 1.134 compared to 1.152) – falling behind the RF and MLP. The absence of a clear winner is by itself instructive: it suggests that PC's four selected features, being physically correlated radiation measurements (Table 4), are not enough of a signal for any of the four models.

With PC-chosen features, all models showed similar performance, with LR yielding an R² of 0.4917, RF giving an R² of 0.5635, MLP yielding an R² of 0.5599, and OptRBF-SVR yielding an R² of 0.4994. No model achieved an R² greater than 0.57 using this set of features, suggesting that the four PC chosen features, which are different radiations that are physically related to each other (see Table 4), were not diverse enough for accurate predictions.

Figure 4 is consistent with Figure 3 in terms of structure, being divided into error metrics (on the left side) and goodness-of-fit metrics (on the right side) in order to prevent combining metrics having opposite meanings. The results obtained by using the feature subset chosen by the RFE algorithm were very similar to those for PC; all the models performed with R² less than 0.60, with the highest value of R² belonging to MLP (R² = 0.5957), slightly exceeding RF (R² = 0.5814), as shown in Table 10. Once again, the Optimized RBF-SVR model performed similarly to Linear Regression (MSE = 1.128 vs. 1.148), showing that neither PC nor the RFE algorithm helped to find a suitable feature subset.

The features selected through RFE gave similar performances across all models, with R² ranging from 0.49 to 0.60, which was in line with PC. The MLP model performed better than other models for these features (R²=0.5957), slightly beating the RF (R²=0.5814). There is no feature subset obtained from either PC or RFE that made any of the five models perform well.

Figure 5 also uses a two-plot design similar to Figures 3 and 4, where the error bars indicate ± one standard deviation calculated over the five repetitions listed in Table 11. Unlike the PC and RFE techniques, the RFFI technique leads to a clear distinction between the models: the Random Forest is the top performer on all measures (R² = 0.896 ± 0.001), followed closely by MLP (R² = 0.889 ± 0.002), the Optimized RBF-SVR (R² = 0.857 ± 0.007), all clearly ahead of Linear Regression (R² = 0.532 ± 0.007). The error bars in the plots for RF and MLP appear visually almost invisible, which proves that the results presented are indeed very consistent over multiple independent runs instead of being dependent on one particular training/testing data split. It should also be noted that the MedAE of the Random Forest is now represented correctly close to zero (0.000 ± 0.000) instead of the erroneous number 4.0 in the earlier single-run pass of the experiment; the number 4.0 was obviously a mistake in transcription (a calculated value of 4.0 × 10⁻⁶ was transcribed as 4.0000).

RFFI’s selected subset of features provided a clear advantage over PC and RFE for all models and was statistically validated in 5 independent repeat executions of the experiment: Random Forest provided the best results (mean R² = 0.8960 ± 0.0008, mean MSE = 0.2295 ± 0.0036), followed by MLP (mean R² = 0.8893 ± 0.0022) and Opt-RBF-SVR (mean R² = 0.8568 ± 0.0072). Low standard deviations in repeats (all less than 0.011 for R²) indicate that results are consistent and not influenced by any lucky train-test set partitioning. This difference between RFFI and PC/RFE is much greater than in the case of any model on PC and RFE, which confirms the selection of physically relevant features (Global Horizontal Radiation, Performance Ratio, Wind Direction, and Tilted Global Radiation; Table 6) by RFFI.

Note: an earlier single-run pass of this experiment reported an anomalous Random Forest MedAE value; across 5 independent repeats with different random samples and train/test splits (reported above), Random Forest's true MedAE is consistently near zero (mean 0.0000, SD 0.0000), confirming the earlier value was not representative and that Random Forest's median error on this feature set is in fact the lowest of all four models evaluated.

Table 12 and Figure 6 show average measured time (± SD in 5 experiments) for training of all models with RFFI feature set (60,000 rows in each experiment's training set). Linear regression was the fastest-trained model (less than 0.01 seconds), although with the worst accuracy performance. The Optimized RBF-SVR was the second-fastest model trained (21.25 ± 1.12 seconds), although faster than the Random Forest (average training time – 35.78 ± 2.42 seconds), with slightly better accuracy. The fastest nonlinear model was MLP (12.76 ± 2.25 seconds) with accuracy near Random Forest. That was a real compromise between accuracy and speed of calculations: RF is the most accurate model, and Optimized RBF-SVR provides the best compromise in terms of accuracy (R² difference near 4%) and speed.

Figure

6

illustrates the exact measured training time for all four models using all three feature selection methods. The training time of the Optimized RBF-SVR varies greatly based on which feature set it uses. It is the slowest among all models when used with PC features (259.8 seconds due to the hyperparameters selected for it in that feature set), of moderate speed with RFE features (45.3 seconds), and relatively fast with RFFI features (21.3 seconds), where it reaches its highest accuracy (Section 4.3). This means that the training time of the Optimized RBF-SVR is not fast in general, but is dependent on the combination of features and their corresponding hyperparameters; therefore, the good balance of accuracy and speed of training achieved with the RFFI feature set (Section 4.4) is more relevant for this research.

The feature selection approach had a major impact on forecasting accuracy. PC and RFE generated feature subsets that limited all models to having an R² value below 0.60, whereas the use of RFFI allowed for R² greater than 0.85 for all three models, RF, MLP, and Optimized RBF-SVR alike. This difference is significantly greater than the one found in previous research, where the same single feature selection approach was used in isolation, showing that the choice of feature selection approach may matter even more than the choice of regression model when performing this kind of forecasting. The benefit of RFFI probably derives from the fact that this approach includes consideration of nonlinear, tree-based feature importance (which includes cyclical variables, such as Wind Direction, where sine/cosine encoding preserves the cyclical nature of the data).

Random Forest had the best performance among the five methods analyzed, marginally beating MLP and Optimized RBF-SVR when trained on RFFI-selected features. All three nonlinear methods showed a significant improvement over Linear Regression, clearly indicating that the photovoltaic power generation cannot be well modeled using a linear relationship between the predictors chosen and the response. However, Optimized RBF-SVR did not have the best raw performance among the five methods but had the second-best performance while taking only about half the time taken by Random Forest to train, hence giving it an advantage wherever training efficiency is a concern (Section 4.4).

To contextualize the proposed framework, Table 13 compares it against prior studies that used the same DKASC dataset and therefore report errors on a comparable measurement scale. It is worth mentioning that the studies to be compared [22][24][26] provide results from other DKASC installations than the one analyzed in this study and, at times, use other measurement scales or normalization approaches. Thus, the results provided in Table 13 should be considered in this context and cannot be used as a strictly apples-to-apples accuracy comparison. Previous works also utilized a greater number of features (7–13) and did not compute any advanced metrics such as MedAE, EVS, and Huber Loss. What makes the current research valuable is not outperforming previous works with higher error values, but demonstrating the possibility of achieving R² greater than 0.85 using a compact set of four features selected with the RFFI algorithm, and using three model types (RF, MLP, and Optimized RBF-SVR).

The obtained results, which have been consistently reproduced in 5 independent repeated experiments with low standard deviations throughout, show that feature selection is significantly more important for the accuracy of predictions of this problem compared to model selection: the set of features selected by RFFI allowed achieving a mean R² above 0.85 for RF, MLP, and Optimized RBF-SVR models, while PC and RFE could not bring any model to R² above 0.60 irrespective of the algorithm. Such a gap is much wider compared to those presented in previous papers analyzing a single feature selection method separately, and it shows how much the choice of feature selection method can be important for this particular forecasting task rather than the choice of a regression algorithm itself. The RFFI's advantage can be attributed to the ability of the method to detect nonlinear and tree-based feature relevance (including the cyclic Wind Direction feature, which has been represented via sine/cosine transform); PC finds linear correlation only and selects four highly correlated radiation measurement features, while RFE, being greedy, selects the Weather Daily Rainfall feature with little variance.