Short-Term Load Forecasting in Distribution Substation Using Autoencoder and Radial Basis Function Neural Networks: A Case Study in India

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

An interesting work, with clear objectives. The main contribution is the attempt to provide an economical forecaster, in terms of input complexity, by employing an autoencoder with static but reliable neural network.

There are some issues that need to be addressed to:

(1) The autoencoder preformed a 33% reduction of the feature vector. Could traditional algorithms like PCA or SVD perform similarly, with a lower computational burden? A comparison analysis would be interesting.

(2) Tables 7 and 8 in the present paper and Tables 7 and 9 from Ref. 26 are the same. Additionally, The authors compare their model to the simple RBFNSS and they refer to Ref. 26 for the training algorithm. Therefore, did the used the simulation experiments that had been conducted in Ref. 26? If this is the case, the authors should clearly mention that the present work is an extension of their previous one.

(3) A more detailed comparison, for example with at least one of the forecasters hosted in Table 1, would highlight the qualities of the proposed scheme.

(4) The caption of the horizontal axis in Fig. 14 is wrong.

(5) Since the authors employed computational intelligence-based forecasting models, the literature review could enriched by adding some recent load predictors based on deep learning and neurofuzzy systems. An indicative list is given below:

1. Abumohse, M.; Owda, A.; Owda, M. Electrical load forecasting using LSTM, GRU, and RNN algorithms. Energies 2023, 16, 2283.
2. Kandilogiannakis, G.; Mastorocostas, P.; Voulodimos, A.; Hilas, C. Short-term load forecasting of the greek power system using a Dynamic Block-Diagonal Fuzzy Neural Network. Energies 2023, 16, 4227.
3. Vanting, N.; Ma, Z.; Jorgensen, B. A scoping review of deep neural networks for electric load forecasting. Energy Inform. 2021, 4, 49.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Autoencoders need at least a reference. Besides, the description in the paper could be improved. Books are always a good reference. This can also be considered for RBFNN.

The multiplier k is in lower case for kV. This is the correct unity in International System.

The unity of the loads (kW) is described only at Figures 10, 11, 12 and 13. The same as for temperature and humidity. None of the tables informs it. Are the loads normalized when using the proposal? I suppose that the losses (errors) are in percent (%).

Why do the authors choose to compare the proposal with Regression Tree?

Observing Figure 14, the error seems to be large, when comparing the actual load and the proposal. In my opinion, this needs a comment.

I think that the authors should improve the comments about the results. Tables and graphics are important, but some comments give credibility to the work.

It is interesting to comment if the web application is already implemented or if it is a proposal.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

In the paper, a machine learning model based on a combination of a radial basis function neural network and an autoencoder is used to forecast the electrical load on a 33/11kV substation located in Godishala, Warangal, India. As mentioned by the authors, one year of historical data on electrical substation and climate is considered to evaluate the effectiveness of the proposed model, this provides a good database.

A very promising subject. However, here are some suggestions.

Why use a radial basis function neural network and an autoencoder represent a better alternative for this type of events?

How could a radial basis function neural network be a differentiator and a contribution?

There are several things that need to be clarified, why was that number of inputs to the neural network chosen? In the literature, several works are already shown related to forecasting the electrical load using different simulation algorithms, including neural networks; In addition, systems are already shown to be embedded and pose a scientific, technological and social contribution, which now new devices regarding the analysis of energy quality represent new challenges for improvement and thus being able to highlight the economic, ecological and social impact that the implementation of the system represents.

Considering that the electrical system is always subject to different changes, how was the generalization of the neural network performed to ensure the accuracy of the electrical load forecast under different conditions?

Is the input data sufficient to maintain the generalization of the network?

Figures 11 and 12 are not relevant, as they do not allow us to evaluate the active power of the proposed methodology in comparison with the current load. Just a lot of points.

Observing figure 14 of the results, could the authors explain why the predicted values ​​of active power by the proposed methodology are so far from the current load? From the seventh hour to the 16th hour? Isn't this a forecast below expectations?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have replied to my comments adequately.

Reviewer 2 Report

Comments and Suggestions for Authors

I think that they have attended the suggestions proposed although the comment above.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors answered the questions satisfactorily.