Short-Term Load Forecasting Method for Industrial Buildings Based on Signal Decomposition and Composite Prediction Model

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you for this interesting research. Some comments to the paper:

It should be highlighted which kind of production area/manufacturing field the industrial building is involved. I assume that the cold load is very depend on the process inside the building. A detailled graphic of the industrial building should be provided. What are the insulation systems of the building? Isn't that influencing the performance?

The proposed framework should be more in detailled explained why this method is chosen? A Table with the computed metrics involved in the simulation will be appreciated.

The literature review should be extended towards what other researchers are investigating in the field of cooling load prediction or energy simulations of industrial buildings. The authors stay silent on that.

It would be beneficial to have comparisons within a proof-of concept. The proof-of concept is missing. A comparison to another real building or with real data would be interesting - presented in a comparison table of predicition and real data.

The graphics are very small (text) and should be shown in a better quality  -especially Figure 10.

 

 

 

 

Comments on the Quality of English Language

There are some spelling and comma errors. Please improve.

Author Response

1.It should be highlighted which kind of production area/manufacturing field the industrial building is involved. I assume that the cold load is very depend on the process inside the building. A detailled graphic of the industrial building should be provided. What are the insulation systems of the building? Isn't that influencing the performance?

Thanks for your valuable comments, we have added descriptions of the manufacturing fields, processes and insulation systems involved in the building, and provided building infographics.

This research collected cold load data for an industrial building in Hebei province, China. The building is mainly used for the manufacturing and assembly of electronic products. It houses heavy industrial production workshops, light industrial assembly areas, and supporting office spaces. Heavy industry production workshops mainly involve high energy consumption processes such as raw material treatment, machining, assembly and welding, and surface treatment, and the cooling load in this area is high in the production stage. The assembly area and office area of light industry focus on precision assembly, testing and quality control of electronic components and small mechanical parts, emphasizing temperature and humidity control and clean air quality to ensure product quality and operator comfort and safety.

The building details are shown in Table 1. Considering the cold winters and hot summers in Hebei Province, the building has also been specially designed with an efficient insulation system, including the use of high-performance insulation materials and double glazing, to reduce energy consumption while maintaining the comfort of the indoor environment. The building design incorporates efficient ventilation and an intelligent air conditioning system to cope with extreme weather fluctuations, ensuring that the cold load remains stable at around 1000 kW to 2000 kW. This guarantees production efficiency and employee comfort.

2.The proposed framework should be more in detailled explained why this method is chosen? A Table with the computed metrics involved in the simulation will be appreciated.

Thanks for your valuable suggestions, we have followed the details in the framework introduction and organized the evaluation indicators into tables.

The proposed cooling load prediction framework provides an efficient and accurate method for cooling load prediction by combining the unique advantages of ISOA, VMD, RF and BiLSTM-Attention. By optimizing VMD parameters, ISOA ensures that the data decomposition process can reveal the inherent characteristics and dynamic change rules of cooling load data, thus improving the basic quality of prediction. Random Forest (RF) is excellent at predicting high-frequency volatility data, capturing instantaneous and complex patterns in the data. BiLSTM-Attention, on the other hand, focuses on capturing long-term dependencies and global trends in data, which works well for low-frequency data. By integrating these techniques, the method can not only handle and predict subtle fluctuations in cooling load data, but also accurately grasp and predict long-term trends, significantly improving the accuracy and robustness of the forecast.

The evaluation index equation and description are shown in Table 2

 

3.The literature review should be extended towards what other researchers are investigating in the field of cooling load prediction or energy simulations of industrial buildings. The authors stay silent on that.

Thanks for your valuable comments, we have added relevant references and added descriptions in the introduction.

• Song Y; Xie H; Zhu Z; et al. Predicting energy consumption of chiller plant using WOA-BiLSTM hybrid prediction model: A case study for a hospital building.Energy Build. 2023, 300: 113642.
• Song C; Yang H; Meng X; et al. A novel deep-learning framework for short-term prediction of cooling load in public buildings. Journal of Cleaner Production. 2024, 434:139796.

 

4.It would be beneficial to have comparisons within a proof-of concept. The proof-of concept is missing. A comparison to another real building or with real data would be interesting - presented in a comparison table of predicition and real data.

Thanks for your valuable comments, we have added a proof of concept section in Chapter 5, as well as a comparison of prediction results for one building.

5.4 Proof of concept of the model

In implementing the proposed prediction model on-site, considering its involvement in complex data processing and computation, a hardware platform with sufficient computational capacity is required to ensure efficient data processing and model operation. This not only guarantees the accuracy of predictions but also facilitates real-time or near-real-time forecasting, providing strong technical support for the establishment of energy consumption management systems in industrial buildings. If the site lacks a robust hardware system, a cloud-edge collaboration approach can be adopted, where data collected on-site is sent to the cloud for processing, with the predictive results then delivered back to the on-site platform, thus reducing costs. Overall, through concept verification and comparison of prediction results, the proposed hybrid prediction method not only demonstrates effective forecasting capabilities but also has manageable hardware requirements for practical application, ensuring smooth deployment and execution in real-world settings, and showing significant potential for practical use.

 

5.The graphics are very small (text) and should be shown in a better quality -especially Figure 10.

Thanks for your valuable comments, we have reset figures 10-12 to improve the clarity and text size

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This study proposes the ISOA-VMD-RF-BiLSTM-Attention method to enhance the accuracy of industrial building cooling load prediction. However, this work should be improved for possible publication. The following considerations should be addressed:

 

1.        The figures and tables in Fig. 1 are unclear. The authors are recommended to reorganize it.

2.        The meanings of many symbols in equations are lacking. The authors should clarify all of them.

3.        In lines 161-162, the authors mentioned that "The SOA is characterized by its simple structure and high flexibility; however, it faces challenges such as susceptibility to local optima and slow convergence speed". Is there any supporting literature? Please provide them.

4.        The comparison between random initialization and Sine chaotic mapping initialization is shown in Figure 2. However, the detailed differences between them are not described in the corresponding texts.

5.        Why did the authors employ BILSTM rather than LSTM?

6.        In Table 1, why are the optimal values 0? Did the authors calculate them correctly?

7.        Figs. 10-12 are too low-resolution.

8.        Line 525: Why is the best model the ISOA-VMD-RF-BiLSTM-Attention model? What are the underlying mechanisms?

9.        In lines 550-551, the authors mentioned, "Future research will explore further improvements, such as optimizing internal hyperparameters of deep learning methods, to enhance the accuracy of cooling load predictions." So, are the results from several models not derived based on the well-trained models with optimized hyperparameters?

10.    What's the critical limitation of this work? Besides, how do we extend the application of this study?

 

 

Comments on the Quality of English Language

No

Author Response

1. The figures and tables in Fig. 1 are unclear. The authors are recommended to reorganize it.

Thanks for your valuable comments, we have restructured Figure 1 to take into account clarity issues.

2. The meanings of many symbols in equations are lacking. The authors should clarify all of them.

Thank you for your valuable suggestions, the meaning of the missing symbols has been added.

3. In lines 161-162, the authors mentioned that "The SOA is characterized by its simple structure and high flexibility; however, it faces challenges such as susceptibility to local optima and slow convergence speed". Is there any supporting literature? Please provide them.

Thanks for your valuable comments, we have added literature support for SOA problems

• Yang B; Li M; Qin R; et al. Extracted power optimization of hybrid wind-wave energy converters array layout via enhanced snake optimizer. Energy. 2024, 293: 130529.
• Yan C; Razmjooy N. Optimal lung cancer detection based on CNN optimized and improved Snake optimization algorithm. Biomedical Signal Processing and Control. 2023, 86:105319.

4. The comparison between random initialization and Sine chaotic mapping initialization is shown in Figure 2. However, the detailed differences between them are not described in the corresponding texts.

Thanks to your valuable comments, we have carried out a detailed analysis of Figure 2 to show the superiority of sinusoidal chaos initialization.

Figure 2 shows the population distribution comparison between random initialization and sinusoidal chaotic mapping initialization in the initial stage of the algorithm. The randomly initialized images show that the distribution of initial solution points in the search space is relatively scattered and concentrated, which may lead to insufficient exploration of the search space. On the contrary, the images of sinusoidal chaotic map initialization reveal a more uniform and widely dispersed population layout covering a larger range of search space, which indicates that sinusoidal chaotic map initialization shows better global exploration potential and more efficient convergence ability.

5. Why did the authors employ BILSTM rather than LSTM?

In this study, we used BiLSTM instead of the traditional LSTM because BiLSTM can simultaneously capture the contextual dependencies of time series data through its bidirectional structure, thus providing more comprehensive contextual information. As a result, BiLSTM has a better performance than LSTM in understanding complex data patterns, improving prediction accuracy and coping with data fluctuations, and is especially suitable for dealing with dynamic and information-rich cooling load forecasting tasks. And we highlighted that in the introduction，references are also added.

6. In Table 1, why are the optimal values 0? Did the authors calculate them correctly?

Thanks for your valuable comment, it is a common case for optimization problems that the optimal value of the test function is zero because it represents the point where the function achieves its minimum value, assuming the function is designed to be minimized. Evaluating the quality of a function or algorithm often involves assessing its ability to approximate zero, as this indicates effective optimization.    During iterations, values that approach zero, often denoted by negative powers of ten, suggest a high level of precision in optimization. Therefore, the closer the values get to zero, the better the optimization performance is considered to be. The references are also attached.

[1] Fateen, S.-E.K.; Bonilla-Petriciolet, A. Intelligent Firefly Algorithm for Global Optimization. In Cuckoo Search Firefly Algorithm; Yang, X.-S., Ed.; Springer International Publishing: Cham, 2014, 315–330.

7. Figs. 10-12 are too low-resolution.

Thanks for your valuable comments, we have reset figures 10-12 to improve the clarity and text size.

8. Line 525: Why is the best model the ISOA-VMD-RF-BiLSTM-Attention model? What are the underlying mechanisms?

Thank you for your valuable comments. The advantage of ISA-VMD-RF-BILSTM-ATTENTION is that the method comprehensively utilizes the advantages of multiple technologies. The ISOA-optimized VMD can effectively decompose the complex cooling load data into multiple frequency components, which improves the clarity and manageability of the data processed by the subsequent model. Random Forest (RF) predictions for high-frequency components can quickly capture short-term fluctuations in data, while BiLSTM-attention's processing of low-frequency components takes advantage of long short-term memory networks to gain insight into long-term trends and complex dependencies through two-way learning and attention mechanisms. The mechanism behind this approach is to complement each other's strengths by combining different technologies to handle both short-term fluctuations in the data and capture long-term trends, thus ensuring the accuracy of the prediction while improving the adaptability and robustness of the model to complex non-linear models.

At the same time, we added the relevant description to the Performance validation of the model.

9. In lines 550-551, the authors mentioned, "Future research will explore further improvements, such as optimizing internal hyperparameters of deep learning methods, to enhance the accuracy of cooling load predictions." So, are the results from several models not derived based on the well-trained models with optimized hyperparameters?

Thank you for your valuable comments.  In the current study, the model results were all tuned by standard hyperparameter optimize methods, yet the potential for enhancement remains.  Hyperparameter optimization constitutes a crucial step in elevating model performance, with the optimal amalgamation of these parameters frequently hinging on the particularities of the data sets and the model’s structural design.  Furthermore, standard tuning methods might harbor more sophisticated schemes for advancement.  Subsequent methodologies could also provide a more comprehensive consideration of the intricate interplay between model performance and hyperparameters.  Hence, in our continued research endeavors, we intend to delve deeper into the realm of hyperparameter optimization techniques to refine our models further.

10. What's the critical limitation of this work? Besides, how do we extend the application of this study?

The principal limitations of this study lie in its data dependency and issues of interpretability. The efficacy of the model may be constrained by specific datasets, which can affect its applicability in broader contexts. Moreover, as a deep learning model, its “black box” nature presents a challenge in understanding the internal logic of its operations, potentially limiting its use in domains where transparency is highly valued. To surmount these limitations, future research needs to explore the use of more diversified data to enhance the model’s generalizability, as well as invest in resources to improve model interpretability, thereby making it more reliable and trustworthy across a wider range of applications.

Author Response File: Author Response.pdf