Optimizing Solar Radiation Prediction with ANN and Explainable AI-Based Feature Selection

Round 1

Reviewer 1 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

Application of XAI methods (LIME and SHAP) not just for model explanation but specifically for guiding feature selection to optimize an ANN for solar radiation prediction. While XAI for feature selection is an emerging area, its detailed application and demonstrated significant improvement in the SR context with ANNs is a valuable contribution. The comparison of multiple ML models provides a good baseline. However, a significant concern for generalization is the very poor performance in the leave-one-station-out cross-validation (Table 4) compared to the 10-fold CV on the mixed dataset (Table 5). This suggests the current model may not generalize well to entirely new, geographically distinct locations without further adaptation or more diverse training data covering wider spatial variations. This aspect needs more prominent discussion. In addition, inconsistencies and ambiguities significantly hampered clarity, particularly regarding feature definitions (input vs. target), data normalization procedures, and some hyperparameter descriptions, as detailed in the following specific comments.

 

Abstract: “This optimized model depicts the power of XAI-based feature selection in enhancing the ANN model for SR prediction.” -> Consider rephrasing for more impact, e.g., “This study demonstrates the significant potential of XAI-driven feature selection to create more efficient and accurate ANN models for SR prediction.”

Line 76: “From this, optimization-based selected features improved the performance of all models compared to individual models.” -> “From this” is vague. Clarify what “this” refers to. Suggestion: “These studies indicate that optimization-based feature selection can improve model performance compared to using all available features.”

Line 90: “he literature also indicates...” -> Typo. Change “he” to “The”.

Line 105-108: “However, we investigate the best-performing model using two XAI techniques: LIME and SHAP, which provide global and local explanations.Unlike previous works, we leverage XAI methods not only to better understand the feature space, but also to identify the most impactful features for SR prediction, which are subsequently used to retrain the best model.” -> The first sentence is somewhat redundant with the preceding sentence. Consider merging or rephrasing. Suggestion: “Building on this, we employ LIME and SHAP not only for global and local explanation of the best-performing model but, distinctively, to guide feature selection for retraining and optimizing this model.”

Line 109: “This approach enhabces...” -> Typo. Change “enhabces” to “enhances”.

Line 121: “Ml models” -> Typo. Change “Ml” to “ML”.

Line 138: “The data was gathered from 41 solar power plants...” -> Table 1 lists 42 unique station names, and Figure 1 caption mentions “# 42 Cite Location”. Please clarify this discrepancy.

Line 141: “K.A. CARE Olaya” -> Table 1 has “K.A.CARE,, Olaya”. Ensure consistency in naming.

Line 144: “The mean and standard deviation value of all datasets are 30.853 and 8.777, respectively provides insight into the variability of the from each site.” -> “value” should be “values”. “variability of the from each site” -> “variability from each site”. Clarify what these mean and SD values (30.853 and 8.777) refer to (e.g., mean and SD of GHI across all records before normalization?).

Line 151: “missing-values” -> “missing values”. Specify which features had missing values and the extent of missingness. Briefly justify the choice of mean imputation.

Line 153-158 & Table 2 & Figure 1: Major confusion regarding normalization.Line 153 states: “each feature is subjected to independent normalization, resulting in a mean of zero and a standard deviation of one.” (This implies StandardScaler). Figure 1 preprocessing step shows: “Standard Scaler for Normalization”. Line 154 states: “All continuous features were normalized using Min-Max scaling to the range [-1, 1].” (This implies MinMaxScaler to a specific range). Line 156 states: “We set the SR values, which are shown by the Global Horizontal Irradiance (GHI) in the range [-1, 1]...”Table 2 shows GHI with Min 0.000 and Max 1.000. Other features have various Min/Max values, some negative.This is contradictory. Please clarify:Which normalization technique was used (StandardScaler or MinMaxScaler)?What was the range for Min-Max scaling if used? Was it [-1,1] for all features, or [0,1] for GHI and [-1,1] for others?Table 2 data: Are these values after normalization? If so, by which method? The GHI values (0 to 1) in Table 2 do not align with a [-1,1] scaling or a mean of 0 and SD of 1 unless 0 and 1 are the scaled min/max. This entire subsection needs to be rewritten for clarity and consistency.

Line 159: “statistic” -> “statistics”.

Line 160: “The final dataset consists of 1,261 records and 21 attributes.” -> Table 2 lists 21 features including “Global Horizontal Irradiance (GHI)”. Is GHI the target variable? If so, there are 20 input features. If GHI is an input feature as well as the target, this is methodologically problematic and needs justification. Please explicitly state the input features and the target variable. The “Remove Class Column” in Figure 1 preprocessing: what was this column? How many features remained before selecting the 21 attributes?

Section 2.1 ML modelsLine 166: “crucial roles” -> “a crucial role”.Line 179: “with with ReLU activation” -> “with ReLU activation”.

Line 188: “As As ML models...” -> “As ML models...”.

Line 189: “poses challengs” -> “poses challenges”.

Line 201: “model 1(x)” -> “model l(x)” (lowercase L for consistency with mathematical notation for functions/models).

Line 215: “p₁ is the SHAP value” -> “φᵢ is the SHAP value” (to match symbol in Eq. 2).

Table 3:LR: “regression coefficient = 1” -> Clarify. For Lasso, this usually refers to the penalty term (lambda or alpha). Is it lambda=1?SVR: “kernel = radial basic function” -> “radial basis function (RBF)”.RF: “impurity =gini” -> Gini impurity is for classification tasks. For regression, common criteria are ‘squared_error’ (MSE) or ‘absolute_error’ (MAE). Please verify and correct; this is a significant issue if Gini was used for regression.BR: “lambda_1 and lambda_2 : v;” -> What does “v” represent? Typo? Clarify these hyperparameter settings. “alpha_2: 1e-6” -> “alpha_2 = 1e-6”.

Line 238: “RMSRE R2, PC” -> “RMSRE, R², PC”.Equation (6) (RMSRE): The denominator is Pᵢ. The text on Line 251 states “normalized by the Oᵢ”. If normalized by Oᵢ, the denominator in Eq (6) should be Oᵢ. Please verify and correct the formula. Using Pᵢ in the denominator can lead to issues if Pᵢ is zero or close to zero.Line 243: “Sį presents prediction errors.” -> Sᵢ is undefined. Please clarify or remove.

Table 4 & Table 5: The performance difference for MLP/ANN (e.g., MAE 2343.47 in Table 4 vs. 0.0093 in Table 5) is extremely large. This needs explicit discussion. Is Table 4 (leave-one-station-out) showing performance in original GHI units (e.g., Wh/m²) while Table 5 (10-fold CV on all data) is on normalized GHI? If GHI is normalized to [0,1], an MAE of 2343 is impossible. This discrepancy is critical and suggests different data scaling or target variable definitions between these experiments. Please clarify units and normalization for reported errors in all tables.

Line 267: “due to limited by its linear nature.” -> “due to its limited linear nature.”Line 268: “improved significantly than LR.” -> “improved significantly compared to LR.”Table 6: “ANN mode” -> “ANN model”.Table 6 Parameter Calculation: The text states the ANN uses 21 input features (derived from (21+1)*64 = 1408 params for the first layer). If GHI (from Table 2) is the target, it should not be an input feature. This implies there are 21 other input features, or GHI is erroneously included as an input. This is a critical methodological point that must be clarified.Line 287: “The model provides a detailed equation for calculating the parameters. [29].” -> Citation [29] (Kachare et al., STEADYNet) is about EEG analysis and seems incorrect for ANN parameter calculation in this context. Please provide an appropriate citation or explain the calculation method.

Line 301: “It is signifying its strong ability...” -> “signifying its strong ability...”

Table 8 & Equation (10): The RC (%) calculation for “lower is better” metrics (MAE, RMSE, MAPE, RMSRE). Eq (10) is (Opt - Base) / Base. For MAE, (0.0024 - 0.0093) / 0.0093 = -74.19%. Table 8 lists 73.66% (positive). Please clarify that for these metrics, the positive RC(%) indicates a percentage reduction or use absolute values and state improvement. The current presentation can be misleading.Line 327: “decrease in MAE and MSE” -> Table 8 shows RMSE, not MSE. Please be consistent.

Line 329-330: “4.879e-06 Floating Point Operational values (FLOPs)” and “6.945e-06”. Table 8 header says “Flops (G)” meaning GigaFlops. 4.897e-06 GFLOPs = 4.897 KFLOPs. This is very low for an ANN, even a small one, but perhaps plausible for a single inference. Clarify if this is per inference and confirm the unit (GFLOPs vs KFLOPs).Line 342: “DHI, DNI, AT, GHIU, and RHU are the most important features...” -> Figure 3 (SHAP) shows DNI, DHI, AT, GHIU, RH as top features. Figure 4 (LIME example) mentions RHU. Which set of 5 features was used for the optimized ANN? Ensure consistency.

List of Abbreviations: ML is listed twice. ΧΑΙ: The manuscript uses “XAI” (Latin X) throughout (title, abstract, text). The abbreviation list uses “ΧΑΙ” (Greek Chi). Standardize to “XAI”.

Line 361: “ANN is the most efficient model” -> “Efficient” can imply computational speed or resource usage. If referring to predictive accuracy, “best-performing” or “most accurate” would be clearer.

Comments on the Quality of English Language

See the specific comments. 

Author Response

 

Subject: Request to Submit a Second Revision of our Manuscript

Dear Editor in chief of Technologies journal,

I hope this message finds you well.

We sincerely thank you and the reviewers for your constructive feedback and valuable suggestions on our manuscript titled " Optimizing Solar Radiation Prediction with ANN and Explainable" (Manuscript ID: technologies-3676064.  We have carefully addressed the comments, made substantial improvements to the manuscript accordingly and highlight the changes in red color.

Thank you once again for your time and support.

 

Comments and Suggestions for Authors

Application of XAI methods (LIME and SHAP) not just for model explanation but specifically for guiding feature selection to optimize an ANN for solar radiation prediction. While XAI for feature selection is an emerging area, its detailed application and demonstrated significant improvement in the SR context with ANNs is a valuable contribution. The comparison of multiple ML models provides a good baseline. However, a significant concern for generalization is the very poor performance in the leave-one-station-out cross-validation (Table 4) compared to the 10-fold CV on the mixed dataset (Table 5). This suggests the current model may not generalize well to entirely new, geographically distinct locations without further adaptation or more diverse training data covering wider spatial variations. This aspect needs more prominent discussion. In addition, inconsistencies and ambiguities significantly hampered clarity, particularly regarding feature definitions (input vs. target), data normalization procedures, and some hyperparameter descriptions, as detailed in the following specific comments.

Abstract: “This optimized model depicts the power of XAI-based feature selection in enhancing the ANN model for SR prediction.” -> Consider rephrasing for more impact, e.g., “This study demonstrates the significant potential of XAI-driven feature selection to create more efficient and accurate ANN models for SR prediction.”

Authors Response: Thank you for your valuable suggestion. The abstract has been revised accordingly. We have rephrased the sentence.

Line 76: “From this, optimization-based selected features improved the performance of all models compared to individual models.” -> “From this” is vague. Clarify what “this” refers to. Suggestion: “These studies indicate that optimization-based feature selection can improve model performance compared to using all available features.”

Authors Response: Thank you for your comment. We have revised the sentence to improve clarity.

Line 90: “he literature also indicates...” -> Typo. Change “he” to “The”.

Authors Response: Thank you for your comment. The typo has been corrected — “he literature” is now updated to “The literature” in Line 90.

Line 105-108: “However, we investigate the best-performing model using two XAI techniques: LIME and SHAP, which provide global and local explanations.Unlike previous works, we leverage XAI methods not only to better understand the feature space, but also to identify the most impactful features for SR prediction, which are subsequently used to retrain the best model.” -> The first sentence is somewhat redundant with the preceding sentence. Consider merging or rephrasing. Suggestion: “Building on this, we employ LIME and SHAP not only for global and local explanation of the best-performing model but, distinctively, to guide feature selection for retraining and optimizing this model.”

Authors Response: Thank you for your valuable suggestion. We have revised the paragraph accordingly to eliminate redundancy and improve clarity.

 

Line 109: “This approach enhabces...” -> Typo. Change “enhabces” to “enhances”.

Authors Response: Thank you for pointing out the typographical error. We have corrected “enhabces” to “enhances” in the revised manuscript.

Line 121: “Ml models” -> Typo. Change “Ml” to “ML”.

Authors Response: Thank you for pointing out the typographical error. We have corrected “Ml” to “ML” in the revised manuscript.

Line 138: “The data was gathered from 41 solar power plants...” -> Table 1 lists 42 unique station names, and Figure 1 caption mentions “# 42 Cite Location”. Please clarify this discrepancy.

Authors Response:  Thank you for highlighting the discrepancy. The dataset originally contains records from 41 solar power plant sites. However, during manuscript preparation, a few site names were inadvertently omitted from Table 1. These missing entries have now been included in the revised version to ensure consistency between the dataset description, Table 1, and Figure 1.

 

Line 141: “K.A. CARE Olaya” -> Table 1 has “K.A.CARE,, Olaya”. Ensure consistency in naming.

Authors Response: Thank you for pointing this out. The naming inconsistency has been corrected in Table 1 to ensure uniformity across the manuscript. The site is now consistently referred to as “K.A. CARE, Olaya” throughout the paper.

Line 144: “The mean and standard deviation value of all datasets are 30.853 and 8.777, respectively provides insight into the variability of the from each site.” -> “value” should be “values”. “variability of the from each site” -> “variability from each site”. Clarify what these mean and SD values (30.853 and 8.777) refer to (e.g., mean and SD of GHI across all records before normalization?).

Authors Response: Thank you for pointing this out. We have corrected the grammatical errors and clarified the context. The revised sentence now specifies that the values refer to the Global Horizontal Irradiance (GHI) across all records before normalization.

Line 151: “missing-values” -> “missing values”. Specify which features had missing values and the extent of missingness. Briefly justify the choice of mean imputation.

Authors Response:  Thank you for the suggestion. We have corrected “missing-values” to “missing values” in the manuscript. A table has been added to present the features with missing values along with their respective counts. Additionally, we have included a brief justification for using mean imputation

Line 153-158 & Table 2 & Figure 1: Major confusion regarding normalization.Line 153 states: “each feature is subjected to independent normalization, resulting in a mean of zero and a standard deviation of one.” (This implies StandardScaler). Figure 1 preprocessing step shows: “Standard Scaler for Normalization”. Line 154 states: “All continuous features were normalized using Min-Max scaling to the range [-1, 1].” (This implies MinMaxScaler to a specific range). Line 156 states: “We set the SR values, which are shown by the Global Horizontal Irradiance (GHI) in the range [-1, 1]...”Table 2 shows GHI with Min 0.000 and Max 1.000. Other features have various Min/Max values, some negative.This is contradictory. Please clarify:Which normalization technique was used (StandardScaler or MinMaxScaler)?What was the range for Min-Max scaling if used? Was it [-1,1] for all features, or [0,1] for GHI and [-1,1] for others?Table 2 data: Are these values after normalization? If so, by which method? The GHI values (0 to 1) in Table 2 do not align with a [-1,1] scaling or a mean of 0 and SD of 1 unless 0 and 1 are the scaled min/max. This entire subsection needs to be rewritten for clarity and consistency.

Authors Response: Thank you for this detailed observation. We apologize for the confusion. We have clarified the preprocessing. A StandardScaler was applied to all continuous input features, and a MinMaxScaler was used only for the target variable (GHI), scaled to the range [0, 1].
The manuscript text has been updated in Section 2 to clearly describe the preprocessing steps.

Line 159: “statistic” -> “statistics”.

Authors Response: Thank you for pointing out the typographical error. We have corrected “statistic” to “statistics” in the revised manuscript.

Line 160: “The final dataset consists of 1,261 records and 21 attributes.” -> Table 2 lists 21 features including “Global Horizontal Irradiance (GHI)”. Is GHI the target variable? If so, there are 20 input features. If GHI is an input feature as well as the target, this is methodologically problematic and needs justification. Please explicitly state the input features and the target variable. The “Remove Class Column” in Figure 1 preprocessing: what was this column? How many features remained before selecting the 21 attributes?

Authors Response: Thank you for your careful observation. We acknowledge the confusion and have clarified the following points in the updated manuscript. In the earlier version of Table 2, one input feature name was missing. This has now been corrected, and the full list of 21 input features plus 1 target feature (GHI) is included. GHI is the target variable used for solar radiation (SR) prediction and not included as an input feature during training. 

The phrase “Remove Class Column” in Figure 1 refers to the removal of an additional class label column originally included for classification tasks in a prior version of the dataset. This column was removed during pre-processing for SR prediction, which is a regression task.

Section 2.1 ML models Line 166: “crucial roles” -> “a crucial role”.Line 179: “with with ReLU activation” -> “with ReLU activation”.

Authors Response: Thank you for pointing out the typographical error. We have corrected this typo error.

Line 188: “As As ML models...” -> “As ML models...”.

Authors Response: Thank you for pointing out the typographical error. We have corrected this typo error.

Line 189: “poses challengs” -> “poses challenges”.

Authors Response: Thank you for pointing out the typographical error. We have corrected these typo errors.

Line 201: “model 1(x)” -> “model l(x)” (lowercase L for consistency with mathematical notation for functions/models).

Authors Response: Thank you for pointing this out. The notation has been corrected from model 1(x) to model l(x) for consistency with standard mathematical function notation throughout the manuscript.

Line 215: “p₁ is the SHAP value” -> “φᵢ is the SHAP value” (to match symbol in Eq. 2).

Authors Response: We appreciate the comment. The symbol p₁ has been replaced with φᵢ in line 215 to align with the notation used in Equation (2) for SHAP values.

Table 3:LR: “regression coefficient = 1” -> Clarify. For Lasso, this usually refers to the penalty term (lambda or alpha). Is it lambda=1?SVR: “kernel = radial basic function” -> “radial basis function (RBF)”.RF: “impurity =gini” -> Gini impurity is for classification tasks. For regression, common criteria are ‘squared_error’ (MSE) or ‘absolute_error’ (MAE). Please verify and correct; this is a significant issue if Gini was used for regression.BR: “lambda_1 and lambda_2 : v;” -> What does “v” represent? Typo? Clarify these hyperparameter settings. “alpha_2: 1e-6” -> “alpha_2 = 1e-6”.

Authors Response: Thank you for your careful review and insightful comments regarding the hyperparameters listed in Table 3. We updated Table 3 and the accompanying descriptions to reflect these corrections for clarity and correctness.

Line 238: “RMSRE R2, PC” -> “RMSRE, R², PC”.Equation (6) (RMSRE): The denominator is Pᵢ. The text on Line 251 states “normalized by the Oᵢ”. If normalized by Oᵢ, the denominator in Eq (6) should be Oᵢ. Please verify and correct the formula. Using Pᵢ in the denominator can lead to issues if Pᵢ is zero or close to zero.Line 243: “Sį presents prediction errors.” -> Sᵢ is undefined. Please clarify or remove.

Authors Response: Thank you for your detailed observation. In Equation (6) by ensuring that the denominator is Oi. Additionally, we have removed the undefined symbol SiS_iSi​ from the sentence on Line 243 to maintain clarity.

 

Table 4 & Table 5: The performance difference for MLP/ANN (e.g., MAE 2343.47 in Table 4 vs. 0.0093 in Table 5) is extremely large. This needs explicit discussion. Is Table 4 (leave-one-station-out) showing performance in original GHI units (e.g., Wh/m²) while Table 5 (10-fold CV on all data) is on normalized GHI? If GHI is normalized to [0,1], an MAE of 2343 is impossible. This discrepancy is critical and suggests different data scaling or target variable definitions between these experiments. Please clarify units and normalization for reported errors in all tables.

Authors Response: Thank you for your observation. We acknowledge the discrepancy and have carefully reviewed and corrected Table 4. The earlier version mistakenly included unnormalized GHI values, leading to inflated error metrics. We have now ensured consistency in target variable scaling across all experiments, including normalization to [0,1], and have updated Table 4 accordingly.

Line 267: “due to limited by its linear nature.” -> “due to its limited linear nature.”Line 268: “improved significantly than LR.” -> “improved significantly compared to LR.”Table 6: “ANN mode” -> “ANN model”.Table 6 Parameter Calculation: The text states the ANN uses 21 input features (derived from (21+1)*64 = 1408 params for the first layer). If GHI (from Table 2) is the target, it should not be an input feature. This implies there are 21 other input features, or GHI is erroneously included as an input. This is a critical methodological point that must be clarified.Line 287: “The model provides a detailed equation for calculating the parameters. [29].” -> Citation [29] (Kachare et al., STEADYNet) is about EEG analysis and seems incorrect for ANN parameter calculation in this context. Please provide an appropriate citation or explain the calculation method.

Authors Response: Thank you for your insightful comments. We have addressed the issues as follows: 1) Thank you for pointing out the typographical error. We have corrected these typo errors.  Initially, the dataset contained 26 features. During preprocessing, we removed non-relevant columns such as Latitude, Longitude, Date, class and also excluded the GHI column as it is the target variable. This results in 21 input features being used in the ANN model, and GHI as the output. The total number of parameters in the ANN is calculated accordingly, based on these 21 input features. We acknowledge that Citation [29] (Kachare et al., STEADYNet) refers to EEG analysis. However, the reference was cited specifically for the method of parameter calculation in fully connected layers, which remains applicable to our ANN architecture.

Line 301: “It is signifying its strong ability...” -> “signifying its strong ability...”

Authors Response: Thank you for your comment, it is changed

Table 8 & Equation (10): The RC (%) calculation for “lower is better” metrics (MAE, RMSE, MAPE, RMSRE). Eq (10) is (Opt - Base) / Base. For MAE, (0.0024 - 0.0093) / 0.0093 = -74.19%. Table 8 lists 73.66% (positive). Please clarify that for these metrics, the positive RC(%) indicates a percentage reduction or use absolute values and state improvement. The current presentation can be misleading. Line 327: “decrease in MAE and MSE” -> Table 8 shows RMSE, not MSE. Please be consistent.

Authors Response: Thank you for your point. In order to address this, we have revised the equation to use the magnitude in the numerator of the equation of 10. We have also corrected the RC (%) calculation errors in the manuscript to ensure consistency.

Line 329-330: “4.879e-06 Floating Point Operational values (FLOPs)” and “6.945e-06”. Table 8 header says “Flops (G)” meaning GigaFlops. 4.897e-06 GFLOPs = 4.897 KFLOPs. This is very low for an ANN, even a small one, but perhaps plausible for a single inference. Clarify if this is per inference and confirm the unit (GFLOPs vs KFLOPs). Line 342: “DHI, DNI, AT, GHIU, and RHU are the most important features...” -> Figure 3 (SHAP) shows DNI, DHI, AT, GHIU, RH as top features. Figure 4 (LIME example) mentions RHU. Which set of 5 features was used for the optimized ANN? Ensure consistency.

Authors Response: The FLOPs values reported in Table 8 represent the number of floating-point operations per inference. We confirm that these values are in KiloFLOPs (KFLOPs), not GigaFLOPs (GFLOPs). The table header has been corrected to reflect this as “FLOPs (K)”. The low values are expected due to the small size of the optimized ANN model. We also confirm that the optimized ANN model was built using the top five features identified through SHAP analysis, namely: DNI, DHI, AT, GHIU, and RH. The mention of “RHU” was a typographical error and has been corrected to “RH” in the revised manuscript.

List of Abbreviations: ML is listed twice. ΧΑΙ: The manuscript uses “XAI” (Latin X) throughout (title, abstract, text). The abbreviation list uses “ΧΑΙ” (Greek Chi). Standardize to “XAI”.

Authors Response: Thank you for pointing out the inconsistency. We have corrected the abbreviation in the abbreviation list by replacing the Greek letter “Χ” (Chi) with the Latin letter “X” to ensure consistency throughout the manuscript.

Line 361: “ANN is the most efficient model” -> “Efficient” can imply computational speed or resource usage. If referring to predictive accuracy, “best-performing” or “most accurate” would be clearer.

Authors Response: Thank you for your comment, it is changed

Author Response File: Author Response.pdf

Reviewer 2 Report (Previous Reviewer 2)

Comments and Suggestions for Authors

This revision has addressed all the reviewers' suggestions.

Author Response

We sincerely appreciate the time and effort the reviewer has taken to evaluate our manuscript. We are pleased to hear that the revisions have satisfactorily addressed all previous concerns. Thank you for your positive feedback on the paper. 

We are grateful for the constructive review process and are glad that our revisions met the expectations. 

Thank you once again for your valuable input.

Round 2

Reviewer 1 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

My comments were addressed. 

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1 General comments.

- The manuscript addresses an interesting problem and shows some results. However, the identified inconsistencies and lack of clarity regarding key methodological details (selected features, parameters, evaluation strategy, metrics) are significant concerns.

- Applying XAI (SHAP/LIME) for feature selection to optimize an ANN for SR prediction is a relevant and useful contribution. Novelty is incremental but solid within the application domain. However, critical inconsistencies regarding ANN parameters, the exact features selected (including the target variable mistakenly listed as input), the RMSRE formula detail, and evaluation methodology undermine confidence.

- Section numbering should start from 1, not 0. Adjust all subsequent section numbers accordingly (1 -> 2, 2 -> 3, 3 -> 4).

- Provide a working URL or accession code for dataset reproducibility.

2. Specific comments

Typos:

- Line 51: “superpassed” -> Typo. Suggest: “surpassed” or “outperformed”.

- Line 68: “outperfomed” -> Typo. Suggest: “outperformed”.

- Line 80: “demostrated” -> Typo. Suggest: “demonstrated”.

- Line 199: “minimizating” -> Typo. Suggest: “minimizing”.

- Line 212: “cordinality” -> Typo. Suggest: “cardinality”.

- Line 267: “conyains” -> Typo. Suggest: “contains”.

 

Section 0 (Should be Section 1) Introduction

- Line 62: “Guermoui et al. reported that Guermoui et al.” → Remove duplicates.

- Line 89: “Training a model on data from one geographical region may not translate well to another...” → Expand with a brief example or citation to support this claim.

Section 1 (Should be Section 2)

- Section 1: Title should be “Materials and Methods”.

- Line 127: “SR is an essential renewable resource...” -> Redundant introductory sentence, consider removing or merging.

- Line 141: “Princess Norah University with six records...” → Justify including sites with minimal data (6 records). Address potential sampling bias.

- Line 146: “The mean solar radiation value of 30.853 indicates the average SR data per site location, and the standard deviation of 8.777 provides insight into the variability...” -> Specify units for these values (e.g., Wh/m² if referring to GHI before normalization). Table 2 shows GHI normalized [0,1], so clarify what these numbers represent.

- Line 157: Table 2 lists negative minima for features like Air Temperature (ATCelsius: -2.546). Clarify if normalization caused negative values and explain the rationale.

- Table 1: “Total 1265” → Discrepancy with Line 158 (“processed dataset has 1,261 records”). And please explain data exclusion criteria (e.g., outliers, missing values).

- Section 1.1 (Should be 2.1): ML models

- Line 173: Bayesian Regression (BR) is mentioned but not adequately described compared to other models like LR, SVR, RF, GBR, ANN. Add a brief description of BR.

- Line 179: “ANN features multiple fully connected hidden layers...” → Specify activation functions for output layers (e.g., linear for regression).

- Line 212: “f(S) is the model’s prediction using input features only...” → Clarify notation: Define “S” as a subset of features and “f(S)” as the model’s prediction using S.

- Section 2: Title should be “Experiments and Results”, not “… results”

Section 2.1 (Should be 3.1): Evaluation metrics

- Line 247-248: “The RMSER calculates the square root of the mean of squared relative errors, normalized by the Oi.” -> The description implies relative error ((O-P)/O). However, Equation (6) shows division by Si: RMSRE = sqrt(mean(((Oi - Pi)/Si)^2)). Define ‘Si’. If it represents the observed value Oi, state this explicitly and ensure consistency between text and formula. If Si is something else (e.g., standard deviation), define it. Suggest clarifying the RMSRE definition and formula.

Section 2.2 (Should be 3.2): Predictive performance

- Table 4: “ANN achieved the lowest MAE of 0.0093...” → Standard deviations (e.g., MAPE: 1.4728 ± 2.3999) exceed means. Discuss implications of high variability.

- Line 252: “Each model is trained using a complete feature set and is 10-fold cross-validated...” -> This describes the initial model comparison.

- Line 264: “The learning rate was set to 0.001...” -> Contradicts Table 3, which lists the ANN learning rate as 0.01. Clarify which learning rate was used.

- Line 268: “The model uses 7,685 parameters...” -> This contradicts Table 5, which lists 2,561 trainable parameters for the ANN with the complete feature set. Furthermore, manual calculation based on Table 5 architecture (Input 21 -> Dense 64 -> Dense 32 -> Dense 1) yields (21+1)*64 + (64+1)*32 + (32+1)*1 = 1408 + 2080 + 33 = 3521 parameters. Please verify the architecture, recalculate the parameters, and ensure consistency between the text and Table 5.

- Lines 271-272: “We use 70% training data and 10% validation data to monitor the model’s generalization ability.” -> This implies a train/validation/test split (presumably 70/10/20 based on Figure 1), which contradicts the 10-fold cross-validation mentioned in Line 252 for model comparison. Clarify the evaluation strategy. Was 10-fold CV used for hyperparameter tuning and initial comparison (Table 4), and then a fixed split used for final ANN training, explanation, and optimization? Specify clearly.

- Line 295: “XAI contributes to this by streamlining...” → Provide evidence (e.g., ablation study) to support claims about reduced training time.

- Line 299: “In this case, selecting only the top 5 features simplified the model structure...” -> Crucial information missing. Explicitly state *which* 5 features were selected based on the XAI analysis (SHAP/LIME). Were they selected based on global importance (SHAP Figure 3) or another criterion?

- Line 327: “GHI, DNI, AT, GHIU, and DNIU are important features selected through XAI techniques...” -> This lists the selected features, but includes GHI (Global Horizontal Irradiance), which is the target variable (output) and should not be an input feature. This is a major inconsistency. Please verify and list the correct top 5 *input* features selected. Based on Figure 3 (SHAP), the top inputs seem to be DHI, DNI, AT, GHIU, and possibly RH or BPU, not DNIU which ranks lower. Resolve this discrepancy between Figure 3 importance and the features listed here.

Section 3 (Should be Section 4): Title should be “Conclusion and Future Works”, not “… future works”

- Definition or clarification of ‘Si’ in the RMSRE formula (Equation 6).

- Explicit statement of the 5 features selected using XAI for the optimized ANN model within the main text body (Section 2.2.1 or 3.2.1).

- Clear and consistent description of the evaluation methodology (distinguish between cross-validation for comparison/tuning and potential train/val/test split for final model training/optimization).

- Correct and consistent reporting of ANN parameters (Line 268 vs. Table 5 vs. calculation).

- Verification and correction of the list of selected features (Line 327), ensuring the target variable (GHI) is not included as an input feature.

- Line 344: “XAI positively impacts the ANN model’s results...” → Overstated. Acknowledge limitations (e.g., dataset size, regional specificity).

 

 

Comments on the Quality of English Language

Lots of typos:

- Line 51: “superpassed” -> Typo. Suggest: “surpassed” or “outperformed”.

- Line 68: “outperfomed” -> Typo. Suggest: “outperformed”.

- Line 80: “demostrated” -> Typo. Suggest: “demonstrated”.

- Line 199: “minimizating” -> Typo. Suggest: “minimizing”.

- Line 212: “cordinality” -> Typo. Suggest: “cardinality”.

- Line 267: “conyains” -> Typo. Suggest: “contains”.

Reviewer 2 Report

Comments and Suggestions for Authors

This paper deals with solar radiation prediction by employing an AI -based method. Six machine learning models are compared using a publicly available dataset and a range of quantitative metrics. Among the evaluated models, the top-performing one is further analyzed using SHAP and LIME, two XAI techniques, to identify and quantify the most influential features. Finally, the selected features are used as inputs to the best-performing model to improve its efficiency and overall performance.

The paper is clearly written, and the topic is interesting. I suggest to the authors to better highlight the novelty, adding a paragraph at the end of the introduction section. 

Reviewer 3 Report

Comments and Suggestions for Authors

1. Whenever a new abbreviation is introduced, its definition must be provided.
2. The manuscript requires thorough linguistic revision.
3. The introduction is logically constructed and appropriately supported by a sufficient number of cited articles relevant to the subject matter. The authors have also incorporated original critical insights regarding solar radiation forecasting methods, based on a review of the current literature.
4. I recommend reversing the order of Table 1 and Figure 1 within the article, as Figure 1 is discussed prior to Table 1 in the text.
5. In Figure 1, “Val” should be replaced with “Validation”, and similar adjustments should be made for other abbreviations. Furthermore, within the ML Model section, abbreviations such as BR and GBR are employed without prior definition—for instance, references to these are only provided on page 6.
6. When discussing the number of records, the authors refer to the figure of 42, whereas Table 2 presents merely 21 features. Consequently, the terms “records” and “features” appear to be used interchangeably, which is incorrect.
7. The data presented in Table 2 raise some concerns; for example, why does the air temperature report a Tmin of –2.546 °C and a Tmax of 1.81 °C, and why are the barometric pressure values given as p_min = –2.654 hPa and p_max = 1.431 hPa? Similarly, various components of irradiance assume negative values at I_min, or at I_max = (1 – 3.089) Wh/m², without any indication of the corresponding time period over which these data are applicable. The same issues appear to affect the other physical parameters visible in Table 2. I assume that similar discrepancies could be expected for the additional parameters not included in Table 2 (which could amount to as many as 42). In the absence of a clear explanation for this issue, it is impossible to properly analyse the presented data and draw valid conclusions.
8. In line 145, the authors state that the mean solar radiation was “30.853”; however, it is unclear what units are being employed, over which period the irradiance was integrated, and whether the value refers to irradiance, insolation, or solar radiation on a horizontal or normal surface. There are too many uncertainties in this instance.
9. The analysis of solar radiation prediction using AI methods cannot be adequately assessed in its current form, as the databases and the parameters contained therein have not been sufficiently well described. There exists an excessive level of ambiguity.

Comments on the Quality of English Language

The English could be improved to more clearly express the research.