Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Brief summary: This paper proposes a demand response system for energy scheduling that takes into account battery degradation effects and costs. The algorithm uses a long short term memory network to forecast demand and electricity prices and reinforcement learning to determine an optimal energy-scheduling strategy to maximise profit and minimise battery health cost. The algorithm is evaluated through a case study which trains the model with data from 1 June to 27 June, 2019 and the trained models are used to forecast demand response data for 28-30 June, 2019. The paper proposes a promising framework for energy scheduling, however the English quality makes it difficult to understand parts of the text, particularly the methodology. In several sections the tone is too informal for a scientific paper. Regarding motivation, the paper makes a good case for its significance, particularly the need to have an optimised DR-based energy scheduling frameworks in microgrids. However, at least in the introduction, they state that current studies on microgrid scenarios rarely consider, which doesn’t seems to be the case. Based on the system model section, their contribution in this topic is on a more realistic model for the battery degradation costs.

 

COMMENTS

Abstract

Substanciate abstract claims/conclusions with numerical results, there was none in the abstract.

Introduction

39 – Briefly explain the main differences between TOU, CPP and RTP.

51- The sentence should be more formal and objective “...researchers hardly take into account.”

80 – Previous

90 – Compared to what others specifically ?

94 – Remainder

Related work

126 – Your claim that RTDR has greater potential to balance supply and demand is based on what ? If possible, add references to substanciate your claim

135 – There is no need to reaffirm the motivation for the paper, that was already done, and is more appropriate, in the introduction

System model

Table 1 is too far from where it was cited in the text

165 – “both purchasing and selling price are”

169 - “...should be strictly higher…”

219 - “Described”

Case Study

While still valuable, the generalisation of conclusions based on a simulation using data for a single month is questionable. If possible, simulate with data for a larger period, at least a handful of months

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Given the significant academic value of research on energy scheduling and battery health assessment learning systems based on sustainable demand response, the authors propose an energy management system for microgrid scenarios and design a battery health monitoring module. Furthermore, based on demand response, a new scheme for Real-Time Pricing (RTP) is put forward, along with a Q-learning algorithm based on reinforcement learning. Specifically, this paper presents the DR-RQL system, which combines Long Short-Term Memory (LSTM) with reinforcement learning to address microgrid uncertainties. It realizes energy scheduling and battery health estimation, optimizes scheduling strategies by predicting electricity prices and demands, improves operational profits, and extends battery life. Empirical results show that it outperforms various benchmark algorithms. However, the paper still has the following shortcomings:

1. The LSTM module in the paper is used for electricity price and demand prediction. Although its performance is evaluated by Mean Absolute Error (MAE), the preprocessing process of input data is not elaborated. The lack of description of this link may affect model reproducibility. It is recommended to supplement the necessary and appropriate data preprocessing steps and parameter settings.

2. The battery health model only emphasizes the impact of cycle aging on battery degradation. Although it is mentioned that the impact of calendar aging is relatively small, no quantitative analysis data of calendar aging is provided, which cannot fully support the conclusion that "calendar aging can be ignored". It is recommended to supplement more robust supporting evidence.

3. The paper points out that DR-RQL improves operational profits by 5.04%-17.31% compared with benchmark algorithms, but it does not split the core sources of profit improvement (such as the contribution of improved LSTM prediction accuracy and the contribution of optimized RQL scheduling strategies). Is it possible to provide or conduct a sub-item attribution analysis of profit improvement?

4. The manuscript does not test the adaptability of the algorithm under different microgrid scales (such as small communities and medium-sized parks). Is it possible to supplement relevant scale scenario experiments to verify the algorithm's universality? If not, it should be indicated in the paper.

5. The research does not specify the hardware environment for model training (such as CPU/GPU model), which affects experimental reproducibility. It is recommended to supplement relevant hardware information.

6. The experiment adopts 48 time slots (30 minutes each) for scheduling, and does not test the impact of different time granularities such as 15 minutes and 1 hour on the accuracy of scheduling decisions (such as the timeliness of capturing peak-valley electricity prices) and system profits. It is recommended to supplement multi-time granularity comparison experiments.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This work proposes a microgrid energy scheduling and battery health management method (DR-RQL) based on demand response (DR) and reinforcement learning (RL), which integrates LSTM for electricity price and load forecasting and introduces a stochastic greedy strategy to enhance exploration efficiency. The research problem is of practical significance, the methodology is well-designed, and the experimental results demonstrate significant performance improvements compared to various baseline methods. However, there are shortcomings in the clear formulation of the problem modeling, comparison with recent related works, detailed description of the experimental setup, and the reproducibility of the code and data. Specific comments are as follows:

1. The authors describe optimization problem P1 as NP-hard and transform it into an MDP, but they do not sufficiently explain why the problem is NP-hard, nor do they provide theoretical justification or prior validation for the design of the state and action spaces in the MDP. Please supplement with relevant explanations or references.
2. Although some related studies are cited, a comprehensive comparison with recent deep reinforcement learning-based methods (such as DDPG and PPO) for microgrid scheduling is lacking. Please include a comparative analysis with these advanced methods to highlight the advantages of the proposed approach.
3. The manuscriptnotes that the LSTM exhibits lag in electricity price forecasting but does not analyze the underlying causes or propose improvement strategies. Have the authors considered incorporating attention mechanisms or more advanced temporal models (e.g., Transformer) to enhance prediction accuracy?
4. Hyperparameters in the reinforcement learning algorithm (e.g., learning rate α=0.001, discount factor γ=0.95) are critical to performance. Please clarify whether these parameters were set via grid search, empirical experience, or references to other literature, and provide a brief sensitivity analysis.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The revised paper has undergone detailed revisions to the review comments. The overall coherence and visual clarity of the paper have been enhanced.