0 Introduction

In recent years, the issue of climate change has attracted more and more attention from all over the world.More than 120 countries have pledged to achieve carbon neutrality in the middle of the 21th century and more than 110 countries have proposed to update the 2030 autonomous contribution target [1-2].China has stated that it will strive to achieve “carbon peak”by 2030 and “carbon neutrality”by 2060, and proposes to build a new power system, which is characterized by high proportion of new energy and power and electrical equipment.All the ideas will provide a Chinese solution for the sustainable development of global power [3].With the steady progress of “carbon peak and carbon neutralization”and construction of new power systems, the proportion of new energy installed capacity continues to increase, driving the flexibility transformation of thermal power units and rapid development of the energy storage industry [4-6].The emergence of many technologies serves to build a multi-energy, complementary, clean, lowcarbon energy system in this period of accelerated energy transformation.

The emergence of the integrated energy system (IES)benefits from development of the energy internet, which is a classic carrier of energy supply and consumption [7].The efficiency of energy utilization is maximized through cascade utilization and multiple complementation of energy.The current research on IES focuses mainly on planning technology, coordinated control technology, and optimal scheduling technology, load forecasting for IES started relatively late.In references [8-10], short-term forecasting of cooling, heating, and electric load was performed based on multivariable phase space reconstruction, Kalman filters,Elman neural networks, and BP neural networks.

Previous studies have mostly used traditional electric load or heating load-forecasting methods.However, it is difficult to improve forecasting results with massive data in a complex environment.Consequently, style="font-size: 1em; text-align: justify; text-indent: 2em; line-height: 1.8em; margin: 0.5em 0em;">Similar days are commonly used in load forecasting and are often associated with clustering methods.Through different perspectives and evaluation strategies, a historical load day that is most similar to the load characteristics of the day to be predicted is selected as a reference to forecast the load of a future period.References [20, 21] improved on the traditional gray correlation analysis method, introducing distance similarity, shape similarity, and a weighted gray correlation degree judgment matrix to select high similarity load days.Reference [22] obtained fine-grained meteorological virtual similar days based on coarse-grained meteorological data and constructed virtual similar load days weighted by the maximum information coefficient(MIC).Current load classification analysis methods are categorized as artificial intelligence neural network methods and cluster analysis methods.Clustering analysis generally uses K-Means clustering, hierarchical clustering, fuzzy clustering, Gaussian mixture model (GMM) clustering, and other methods.Both hierarchical clustering and partition clustering use hard partitioning, with each object to be identified strictly classified in a certain category; the eitheror nature is not suitable for large-scale datasets.GMMs allocate cluster members according to the cluster probability,which is a soft classification.Classification using GMM does not place each object in a specific cluster, but allocates the probability or possibility of the object in each cluster.Thus, it can effectively avoid hard allocation, and provides more information than hard allocation clustering [23-26].

These previous studies focused on selection of similar days for a single load using relatively simple methods.The daily characteristics of the load were mostly rough meteorological characteristics or single-load characteristics;there was a lack of more refined processing in construction of the feature sequence, which could not fully reflect the essential law of the user energy consumption behavior on a given load day.There were also insufficient data types and data volume.Most of the literature on selection of similar days is limited to certain historical load day data(static selection), with many contingencies in using it as input.Selecting similar days requires global evaluation of the multi-dimensional and multi-view features in IES when the load pattern becomes complex and the coupling degree increases.

Although the traditional single-load method is relatively mature, joint-load forecasting of IES is still in its infancy.Due to the complex coupling characteristics of multiple loads, there are few improvements and re-applications of traditional load-forecasting methods in current IES loadforecasting research.Instead, a deep learning algorithm is used for pre-feature processing and subsequent prediction to realize “high-dimensional”feature mining of the load.However, use of the dominant features of the load is still lacking.Unlike single-load forecasting, IES includes different multi-loads such as electric load, heating load,and cooling load, which can more comprehensively and multi-dimensionally display user energy consumption rules;the sensitivity of the loads to meteorological data is also different.Based on previous research, a load-forecasting method for IES based on LSTM and dynamic similar days with multi-features is proposed in this paper.The main contributions are summarized as follows:

(1) A new generalized load day sequence with multiple characteristics is constructed using the differences in time scale, load type, and meteorological data, covering the diversity of load type, the comprehensiveness of meteorological characteristics, and the difference in time granularity (fine-grained characteristics play a major role and coarse-grained characteristics play an auxiliary role).The reconstruction of an IES comprehensive load day with distinct characteristics makes the description of any load day more comprehensive and three-dimensional, and solves the problem of limited information in single-load forecasting.

(2) The typical scenes of a comprehensive load day are divided using the advantages of soft classification in the GMM method.Combined with the gray correlation analysis method, the similar day set for the day to be predicted is found in the typical scene and a dynamic similar day is calculated according to the weight.It covers the comprehensive characteristics of multiple similar days and removes the “static similar day”constraint in previous research.

(3) With data from dynamic similar days and key information from adjacent load days, the advantages of deep learning in prediction are effectively utilized.The LSTM method is used to capture the time correlation of feature sequences in high-dimensional space to achieve highprecision joint-load forecasting of the IES.

The remainder of this paper is organized as follows.The multi-morphological characteristics of an IES load day are analyzed in Section 2.Section 3 introduces the proposed dynamic similar day construction method based on load-scene clustering using a GMM and gray correlation analysis.Section 4 introduces the framework of the IES load-forecasting method based on the dynamic similar day and LSTM.In Section 5, the proposed method is verified using an example.Section 6 presents the conclusions and objectively explains the limitations of the method in terms of generalization and robustness, and presents a future research direction.

1 Multi-morphological characteristics of IES load days

Historical load-day characteristics are often used in forecasting models to improve accuracy in short-term or ultra-short-term load forecasting.Most studies use periodic characteristics as a prediction reference, such as adjacent time, adjacent day, and adjacent week in forecasting,although the reference value may only be reflected in the similarity of a single form.From the perspective of the macro change of the load, a similar day of load is simply the change rule of the final load, which is a comprehensive presentation of multiple characteristics and is inseparable from other factors such as meteorology.To make the selection method of IES similar load days more robust and extend the depth of feature dimensions and the width of contained information, the combined sequence of coarse and fine time granularity is used to describe the complete load day characteristics, as shown in Table 1.

Considering the cooling, heating, electric, and meteorological data of an IES park in northern China as the object, the Pearson correlation coefficients of any load day with continuous adjacent days and non-adjacent days were calculated according to the heating load, cooling load,electric load, day-ahead feature, intra-day feature, and full feature sequence.The Pearson correlation coefficient is calculated as

where ^X and Y^ are any two sequences; xˆ and yˆ are individuals in sequences ^X and Y^, respectively, and ψ is their sum.Usually, a positive ρ indicates a positive correlation, and a negative ρ indicates a negative correlation.The degree of correlation is judged according to absolute value.The correlation coefficient heat map is shown in Fig.1.

Table 1 Characteristic composition of comprehensive load day

Fig.1 Heat map of Pearson correlation coefficient between comprehensive load days

A load day with feature expansion shows two basic characteristics from the degree of color depth change:

(1) The characteristics of comprehensive load days consist of multivariate dynamic load and meteorological data of different granularities, including characteristics of adjacent days and of the same day in adjacent weeks.The feature form is rich; the color consistency of the correlation coefficient heat map in the adjacent period is strong and the feature change is stable.

(2) There is no significant difference or boundary between the morphological characteristics of a load day with a single load as an independent feature in adjacent and non-adjacent periods; the discrimination is not strong and has a certain randomness.In contrast, under the influence of meteorological characteristics with different granularities and time scales, the similarity between long-period interval load days is significantly different, which refines and highlights the difference in load in the time dimension.

2 Load-scene clustering and dynamic similar day construction

The key to short-term load forecasting is using known information to mine unknown rules.In the structure of the extended load day, the fine-grained features are hourly electric, heating, and cooling load, and the coarse-grained features are the meteorological characteristics of the day,and the load and meteorological characteristics of the adjacent day and adjacent week.The historical load days are clustered according to the full-load characteristics using the GMM method, and the typical scene change rules are summarized.Based on the gray correlation analysis, the coarse-grained average feature change curve is used to match the scene of the forecast day.Five load days with the highest correlation degree are mined in the scene; their finegrained load characteristics are weighted to construct the key data input for the forecasting model.

2.1 Basic principles of GMM

The GMM clustering method is used to divide the scene, usually using a combination of multiple Gaussian distribution functions to estimate the sample distribution;the purpose of clustering is to maximize the probability density of the sample.The GMM can be defined as

The GMM outputs a series of probability values after the classification training; n Gaussian distribution functions correspond to n categories.The final clustering result of the sample depends on the corresponding maximum probability value.To determine the optimal number of clusters, most studies use the Bayesian information criterion (BIC) to measure the clustering effect of the GMM; the standard is the lowest BIC value [27, 28].To solve the maximum likelihood estimation of model distribution parameters,the expectation-maximization algorithm (EM) is used to estimate GMM parameters [29].

2.2 Gray correlation analysis

(1) Feature expansion: Based on the coarse and fine time granularity and the far and near time periods, the daily load characteristics in the dataset are expanded to construct a 153×1 comprehensive daily load sequence.

(2) Scene division: The GMM method is used to cluster the comprehensive load days with expanded features to form typical scenes with key IES features.The division of any scene is a comprehensive result considering the current and historical load and the law of meteorological changes.

(3) Scene matching: The coarse-grained features are supplemented using the feature expansion method for any load day to be predicted (future hourly electric, heating,and cooling loads are unknown).The coarse-grained meteorological data of the day uses the weather forecast value; other coarse-grained data are derived from historical values.An 81×1 matching sequence is formed and the gray correlation degree is calculated with the average coarsegrained sequence of the divided scenes.The scene with the highest degree of correlation is selected as the set of similar day scenes.

(4) Dynamic selection: The gray correlation degree of the coarse-grained sequence is calculated with all load days in the similar day scene, and the five load days with the highest correlation degree are selected.Assuming that their values are σi(i=1,2,3,4,5), their respective information importance is calculated as

3 IES load-forecasting method based on dynamic similar days and LSTM

3.1 Long short-term memory

LSTM is an optimized structure of a recurrent neural network (RNN) for improving gradient explosion and gradient dispersion problems, with a good processing effect on time-series problems.A standard LSTM consists of an input gate, forgetting gate, output gate, and storage unit state.The LSTM process at time step t can be summarized as

Fig.2 Selection method for dynamic similar days

In the formula, it, f t, ot, gt, and ctare the input gate,forgetting gate, output gate, input modulation gate, and storage cell state, respectively; σ is the sigmoid function; ⊙is the product symbol; φ is the hyperbolic tangent function;W*x and W*h are weight matrices; b* is the bias vector.

3.2 Load-forecasting framework based on dynamic similar days and LSTM

This study uses the similar day method to improve the accuracy of IES load-forecasting with multiple features.The method includes feature construction, scene division,scene matching, and weighted calculation.The dynamic similar day for the day to be predicted is obtained, and the explicit information input in the prediction process is increased.We use LSTM to capture the time correlation of feature sequences in high-dimensional space.The prediction process includes three stages, as shown in Fig.3: scene division of the comprehensive load day, dynamic similar day construction, and time-dependence capture.

Fig.3 The prediction process

Scene division of comprehensive load day: According to the IES load characteristics and information importance,a comprehensive load day with load information and meteorological information is constructed, including near and far time periods, and coarse and fine time granularity.The GMM is used to divide the typical scenes of historical load days to obtain the typical IES load day scenes.

Dynamic similar day construction: The coarse-grained characteristics of the forecasting day are supplemented using the comprehensive load day construction method, and the gray correlation degree is calculated with the average coarse-grained sequence of the divided scenes to determine the best scene.The gray correlation degree of the coarsegrained sequence is calculated with any load day in the best scene.Five load days with the highest correlation degree are selected to calculate the dynamic similar day.

Time-dependence capture: Using the dynamic similar day and other information as input, the time correlation contained in the feature sequence is extracted by LSTM in high-dimensional space for prediction of the IES multivariate load and to improve prediction accuracy.

The specific input features of LSTM are described as follows.Thirty key factors are selected as input characteristics at any time t in the LSTM input layer,including multivariate historical load data, temperature,humidity, and illumination radiance.The specific reference time points are time t of the adjacent day and dynamic similar day.The load data include three parts: hourly data for electric, cooling, and heating load within 24 h, average data within 4 h, and average data within 6 h.In addition to the hourly data of time t for adjacent days and dynamic similar days, the temperature, humidity, and light radiance also include forecasting data for the current day (forecasting results from two weather forecasting institutions).The input characteristics at time t can be expressed as

In the formula, L is the adjacent day data, D is the dynamic similar day data, Q is the weather forecasting data of the day, and subscripts 1 and 2 represent the number of the forecasting institution; e, c, and h represent electric,cooling, and heating load, respectively; e/4, c/4, and h/4 are the average loads within 4 h, and e/6, c/6, and h/6 are the average loads within 6 h; w, s, and f are the temperature,humidity, and irradiance; w/y, s/y, and f/y are their forecasting values.

The IES load-forecasting model based on LSTM and dynamic similar days with multi-features is shown in Fig.4.

Fig.4 IES load-forecasting model based on LSTM and dynamic similar days with multi-features

4 Analysis of examples

The multi-load data required in this study are at least more than two consecutive years and the time resolution is not less than 1-h, so there are higher requirements for data quality and quantity.To the best knowledge of the authors,there are mainly five types of datasets used in the current research on IES load forecasting, as listed in Table 2.

These five types of datasets mainly have the following characteristics: datasets of industrial park and comprehensive district are almost not open to the public;datasets of residential buildings and commercial buildings can hardly meet the requirements of the amount of data.Therefore, campus dataset has become the best choice forthis study.To our knowledge, the cooling, heating, and electric load data at different time scales from Arizona State University in the US have been widely used in recent research, which are open to the public and have the advantages of high data quality and complete data types compared with other datasets.

Table 2 Different types of IES datasets

In this study, load data were from a large building IES at Arizona State University Tempe campus and meteorological data were derived from the official website of the National Renewable Energy Laboratory, and all the resolution is 1-h.Tempe campus has nearly 300 buildings,the climate is characterized by hot summers, warm winters,and abundant sunshine throughout the year.The basic data were recorded in the campus Metabolism Project network platform at the school, which is an interactive web tool that enables the user to view the current resource use on campus.Campus Metabolism was created to be used for learning,teaching, researching and empowering change.Constructing a similar day set using historical data from 2017-2018,and the prediction model was trained and verified using the historical data from January 2018 to August 2020; the ratio of the training set to the test set was 7:3.Historical feature data were normalized to facilitate model training and comprehensive load day visualization.The normalization method is expressed as

For the LSTM hyperparameters, the results show that setting three layers of LSTM produces better results.The number of neurons in each layer was 128, 128, and 64; the LSTM time step parameter (Timesteps) and the number of neurons in the fully connected layer were both 24.

4.1 Analysis of clustering results

The 2017-2018 IES load days in the region were clustered into eight typical scenes after normalization, as shown in Fig.5.

Fig.5 Eight typical classification scenes

In Fig.5, the first half of the curve is relatively smooth, representing the fine time granularity multivariate load value.The second half represents the coarse time granularity value for the intersection of weather and load.The load trends in different scenes are generally different from the overall clustering results, and there are significant differences in the size and trend of key loads and meteorological values for the adjacent days and weeks.

The eight GMM scenes can be divided into low load demand (Scene 1, Scene 3, Scene 5), medium load demand(Scene 6, Scene 7, Scene 8), and high load demand (Scene 2, Scene 4) according to the load demand.The minimum and maximum electric load and cooling load are almost coincident and exhibit the opposite trend of the maximum heating load.Scene 1, Scene 3, and Scene 5 reflect the slight change in heating load as the ambient temperature decreases; the heating load and ambient temperature play a significant role in the comprehensive features in the classification.With a continuous increase in temperature in Scene 6, Scene 7, and Scene 8, the thermal load exhibited an irregular low level and the electric load was not changed significantly.The significant increase in cooling load became the key classification factor.The climate gradually changed from dry to wet, which also became a sensitive classification feature.Scene 2 and Scene 4 are in the summer (high-temperature period); thus, the heating load is the lowest and changes irregularly.The electric load and cooling load are the highest, and the change characteristics are significant.During this period, the solar radiation,temperature, and humidity are also at the highest levels for the year.

On the whole, the results of GMM classification were generally reasonable.With multiple features, sensitive features of the comprehensive load day can be considered to achieve the optimal division.Compared with other hard classification methods, with the GMM method we can more intuitively observe possibilities in the classification process for any individual.

4.2 Method construction and result comparison

4.2.1 Comparison between dynamic similar days and fixed input

To compare the impact of dynamic similar days on the prediction results, fixed inputs are used:

Method 1: Related data for the dynamic similar day in the input is removed and retrained.

Method 2: Retrain with data from two days before instead of dynamic similar days.

Table 3 Forecasting results obtained by different methods

The overall prediction of method 2 is better than that of method 1 with more perfect data input, but the improvement in prediction accuracy is limited by the test set results.When adjacent days or two consecutive adjacent days are selected as input, the accuracy of load-forecasting is considerable in a partially continuous period.This method has two conditions in practice: (1) the characteristics of continuous load days are quite different, resulting in a completely different load trend; (2) some discontinuous load days with similar daily characteristics cannot provide effective information as input to improve prediction accuracy due to the influence of a fixed model.Although the performance of the proposed forecasting model is not as good as that of method 1 and 2 at some points, the overall MAPE value is increased by 0.7%-1.5%on average, and the RMSE value is lower, demonstrating that the overall prediction of the model is relatively stable in a certain range and is more robust based on the test set results.The accuracy comparison for a continuous 20 days is shown in Fig.6.

Fig.6 Accuracy comparison for a continuous 20 days

4.2.2 Dynamic similar day for scene without feature expansion

The comprehensive load day constructed in this study is a multi-dimensional description of a load day, dominated by load characteristics and supplemented by meteorological characteristics, covering the current time state and historical state.To further illustrate the impact of the extended features on the prediction results, we use only the electric, cooling,and heating loads for GMM clustering and subsequent prediction.In this case, the number of features for each load day is reduced from the original case by 68 (48 load values on the day, 12 (8+4) load values on the adjacent day, and 8 load values in the adjacent week).The number of features is reduced by 20 (12 (8+4) load values on the adjacent day and 8 load values in the adjacent week) in the subsequent matching of similar days.The final input dimension is 11 ×24.The test results are shown in Table 4.

Table 4 Test set results of different models

Compared with the proposed method, the prediction of a scene without feature expansion is relatively poor, with almost no advantage in prediction accuracy (MAPE) or stability (RMSE).The comprehensive load day with extended multi-load characteristics exhibits completely different characteristics from the single-feature scene in GMM classification.Considering the electric-load comparison as an example, a continuous 120 days are randomly selected; the classification results are shown in Fig.7.

Fig.7 Classification results for randomly selected continuous 120 days

Fig.8 Forecasting load for five consecutive days

It is observed that a comprehensive load day with expanded multi-load characteristics has stronger continuity in GMM classification and more stable classification results.The proportion of unstable classification in all sets (not in the same classification set for two consecutive days) was calculated; the proportion for single scenes was 53.6% and the proportion for extended scenes was 22.5%.Compared with the actual operation experience, the overall reference value of a continuous load day is larger, whereas single-load clustering results show more randomness and contingency.Although target features are more similar in the same category, there are also more limitations in predicting unknown variables.The forecasting load for five consecutive days (numbers 1-5) is shown in Fig.8.

Considering the fourth day as an example, the classification method based on the electric load (single-load)classifies it to the set of Days 1-3.However, the electric load changes suddenly on the fourth day and exhibits completely different characteristics from the set, which results in a large deviation of the predicted value.Under the influence of cooling and heating load trends, the proposed method classifies it into a set corresponding to Day 5, with significantly improved prediction accuracy.It is observed that dynamic similar days without feature expansion scenes only consider the independent morphological trend centered on their respective loads, whereas dynamic similar days with feature expansion scenes have a broader perspective,which aims at mining wider coverage and deeper levels of feature dynamics to obtain more effective information for the forecasting model.

4.2.3 Differences in prediction results of test sets in different scenes

The multi-dimensional information of the integrated load form in IES can describe user side behavior more accurately, unlike single-load forecasting of independent power systems or thermal systems.The prediction results for the test set are shown in Table 5.Some basic laws can be obtained by mining the cooling, heating, and electric load relationships in Scene 8 of GMM classification.

The overall prediction accuracy of Scenes 1, 3, and 8 is higher; the prediction accuracy of Scenes 4 and 6 is relatively poor using the proposed method.Ten load days were randomly selected in each scene to calculate the gray correlation degree of electric, heating, and cooling load with other load days in the scene.Compared with Scene 3 and Scene 4, the correlation values of electric and cooling load,electric and heating load, and cooling and heating load were[0.60, 0.39, 0.68] and [0.53, 0.04, 0.12], respectively.The average change curve for the six key features in Scenes 3 and 4 is shown in Fig.9.

From Fig.9, Scenes 3 and 4 have typical load divisions.In Scene 3, the average temperature and solar radiation are low and the heating load demand is high, whereasthe electric and cooling loads are low.The electric load and cooling load in Scene 4 increase significantly and the heating load supply remains stable and low with the arrival of summer, almost a straight line.Overall, the electric load and cooling load base are large and the change rule is relatively fixed, whereas the heating load base is small and shows completely different fluctuation rules in different seasons.Thus, although the relative values of electric load and cooling load in Scene 3 are small, its fluctuation law is still similar to that in Scene 4; the correlation coefficients between electric load and cooling load in the two scenes are almost the same.However, the heating load shows a completely different condition in Scene 4.The change of heating load does not show a fixed rule, which leads to a significant decrease in the correlation degree of electricheating load and cooling-heating load compared with Scene 3.In the subsequent prediction, a heating load with a low correlation degree does not produce effective information,and adds irrelevant noise to the model, resulting in a lower overall prediction accuracy of multivariate load than in other scenes.

Table 5 Prediction results for test sets in different scenes

Fig.9 Average change curve for six key features in Scenes 3 and 4

4.2.4 Comparison of forecasting results based on other clustering methods

To demonstrate that the GMM in this study has advantages over other methods in scene clustering,K-means clustering, fuzzy C-means clustering (FCM),and hierarchical clustering methods were used to divide the scene and construct dynamic similar days.LSTM was used to forecast multiple loads.The indicators for different forecasting methods are shown in Table 6.

Table 6 Forecasting results with different clustering methods

From the results, the optimal clustering number for K-means, FCM, and hierarchical clustering is less than that for GMM, which means that the GMM classification results are more detailed and carry more information.The refined classification results offer more possibilities to mine the correlation of multiple comprehensive features.The MAPE and RMSE for GMM are lower than for the other methods.Scene division based on GMM provides a higher matching degree for construction of dynamic similar days, indicating certain advantages in prediction accuracy and stability.

5 Conclusion

To obtain the characteristics of complex energy coupling and rich feature information of IES, a load-forecasting method for IES based on LSTM and dynamic similar days with multi-features was proposed.Using GMM and gray relational analysis to construct multiple dynamic-load similar days, LSTM was used to improve the prediction accuracy.The main conclusions of this paper are presented as follows:

(1) The comprehensive load day based on the intersection of coarse and fine time granularity and far and near time periods has both multiple-load forms and meteorological forms, and can reflect the multi-dimensional characteristics of load days in a deeper and more stable manner.

(2) Selection of dynamic similar days is based on GMM and gray relational analysis, which use coarse-grained feature matching and fine-grained feature weighting to dynamically and effectively update the input.

(3) Methods considering multiple-load information have better performance than methods based on single-load characteristics.

There are some limitations in the method used in this study.

(1) The feature construction of coarse and fine time granularity and the feature sequence construction of far and near time periods mostly use empirical results.It is necessary to further explore which features are selected and realize adaptive construction of feature sequences to maximize the performance of prominent features without losing basic features, and to make the characteristics of any load day more obvious.

(2) Combination with other deep learning algorithms such as convolutional neural networks has been rare; based on the results of previous research, further exploration is needed.There is still room for improvement in the exploration, fusion, and extraction of the high-dimensional feature space of load characteristics, especially in the early stage, to further improve current prediction accuracy.

IES currently plays a vital role in the energy revolution and transformation.Dynamically updating the input and forecasting model framework to maximize the use of key multivariate feature information should be a follow-up research direction.

Acknowledgements

This work is supported by National Natural Science Foundation of China (NSFC)(62103126).

Declaration of Competing Interest

We declare that we have no conflicts of interest.

References

[1] Han X Q, Li T J, Zhang D X, et al.(2021) New Issues and Key Technologies of New Power System Planning Under Double Carbon Goals.High Voltage Engineering, 47(9):3036-3046 (in Chinese)

[2] Xiao X Y, Zheng X X (2022) New Power Systems Dominated by Renewable Energy Towards the Goal of Emission Peak &Carbon Neutrality:Contribution, Key Techniques, and Challenges.Advanced Engineering Sciences, 54(1): 47-59(in Chinese)

[3] Xie X R, Ma N J, Liu W, et al.(2023) Functions of Energy Storage in Renewable Energy Dominated Power Systems:Review and Prospect, Proceedings of the CSEE, 43(1):158-168(in Chinese)

[4] Chang Y K, Liu R, Ba Y, et al.(2019) Frequency Control Mode and Performance Assessment in UHV grid with high proportion.Power System Technology, 43(2):621-631(in Chinese)

[5] Huang Y H, Ding T, Li Y T', et al.(2021) Decarbonization technologies and inspirations for the development of novel power systems in the context of carbon neutrality.Proceedings of the CSEE, 41(S1):28-51(in Chinese)

[6] Zhuo Z Y, Zhang N, Xie X R, et al.(2021) Key technologies and developing challenges of power system with high proportion of renewable energy.Automation of Electric Power Systems,45(9):171-191(in Chinese)

[7] Wu J, Yan J Y, Jia H J, et al.(2016) Integrated energy systems.Applied Energy, 167: 155-157

[8] Guan C, Luh P B, Michel L D, et al.(2013) Hybrid Kalman filters for very short-term load forecasting and prediction interval estimation.IEEE Transactions on Power Systems, 28(4): 3806-3817

[9] Liang R, Wang H T, Wu K H, et al.(2020) Short- term forecasting of cooling, heating and power loads based on neural network and ARIMA model.Proceedings of the CSU.EPSA,32(03): 52-58(in Chinese)

[10] Ma D Y, Sun B, Liu C(2021) Short-term cooling and heating power load prediction method based on multi-weather information.Power System Technology, 45(03):1015-1022(in Chinese)

[11] Zhu J Z, Dong H J, Li S L, et al.(2021) Review of name="ref12" style="font-size: 1em; text-align: justify; text-indent: 2em; line-height: 1.8em; margin: 0.5em 0em;">[12] Li R, Sun F, Ding X, et al.(2020) Ultra Short-term load forecasting method for user-level integrated energy system considering spatio-temporal coupling of multi-energy, Power System Technology, 44(11):4121-4134(in Chinese)[13 Wu K H, Wu J, Feng L, et al.(2021) An attention-based CNNLSTM-BiLSTM model for short-term electric load forecasting in integrated energy system.International Transactions on Electrical Energy Systems, 31(1): e12637

[14] Zhu R J, Guo W L, Gong X J (2019) Short-term load forecasting for CCHP systems considering the correlation between heating, gas and electrical loads based on deep learning.Energies, 12(17): 3308

[15] Wang B, Zhang L M, Ma H R, et al.(2019) Parallel LSTMbased regional integrated energy system multienergy source-load information interactive energy prediction.Complexity, 2019: 1-13

[16] Wang X, Wang S X, Zhao Q Y, et al.(2021) A multi-energy load prediction model based on deep multi-task learning and ensemble approach for regional integrated energy systems.International Journal of Electrical Power & Energy Systems, 126: 106583

[17] Zhao H R, Guo S (2021) Uncertain interval forecasting for combined electricity-heat-cooling-gas loads in the integrated energy system based on multi-task learning and multi-kernel extreme learning machine.Mathematics, 9(14): 1645

[18] Sun Q K, Wang X J, Zhang Y Z, et al.(2021) Multiple load prediction of integrated energy system based on long short-term memory and multi-task learning.Automation of Electric Power Systems, 45(05):63-70(in Chinese)

[19] Shi J Q, Tan T, Guo J, et al.(2018) Multi-task learning based on deep architecture for various types of load forecasting in regional energy system integration.Power System Technology,42(03):698-707(in Chinese)

[20] Wu Y, Lei J, Bao L, et al.(2018) Short-term load forecasting based on improved grey relational analysis and neural network optimized by bat algorithm.Dianli Xitong Zidonghua/Automation of Electric Power Systems, 42(20): 67-72

[21] Wu X Y, He J H, Zhang P, et al.(2015) Power system shortterm load forecasting based on improved random forest with grey relation projection.Automation of Electric Power Systems,39(12):50-55(in Chinese)

[22] Li B, Gao F(2021) Short-term load forecasting for regional power grid based on virtual similar days and dual-stage attention-based long and short Time pattern network.Automation of Electric Power Systems, 45(22):55-64(in Chinese)

[23] Lin S F, Tian E W, Fu Yang (2017) A load classification method based on information entropy segmentation aggregation approximation and apectral clustering.Proceedings of the CSEE,37(8):2242-2253(in Chinese)

[24] Li L L, Sun J, Wang C H, et al.(2019) Enhanced Gaussian process mixture model for short-term electric load forecasting.Information Sciences, 477: 386-398

[25] Debnath R, Bardhan R, Misra A, et al.(2022) Lockdown impacts on residential electricity demand in India: A name="ref26" style="font-size: 1em; text-align: justify; text-indent: 2em; line-height: 1.8em; margin: 0.5em 0em;">[26] Zhang Y F, Ai Q, Li Z Y (2020) Improving aggregated baseline load estimation by Gaussian mixture model.Energy Reports, 6:1221-1225

[27] Mirzal A (2022) Statistical analysis of microarray data clustering using NMF, spectral clustering, kmeans, and GMM.IEEE/ACM Transactions on Computational Biology and Bioinformatics,19(2): 1173-1192

[28] Shi J Y, He Q J, Wang Z L (2019) GMM clustering-based decision trees considering fault rate and cluster validity for analog circuit fault diagnosis.IEEE Access, 7: 140637-140650

[29] Shu X B, Zhang L Y, Sun Y L, et al.(2021) Host-parasite:Graph LSTM-in-LSTM for group activity recognition.IEEE Transactions on Neural Networks and Learning Systems, 32(2):663-674

[30] Powell K M, Sriprasad A, Cole W J, et al.(2014) Heating,cooling, and electrical load forecasting for a large-scale district energy system.Energy, 74: 877-885

[31] Wang X, Wang S X, Zhao Q Y, et al.(2021) A multi-energy load prediction model based on deep multi-task learning and ensemble approach for regional integrated energy systems.International Journal of Electrical Power & Energy Systems, 126: 106583

[32] Wang S M, Wang S X, Chen H W, et al.(2020) Multienergy load forecasting for regional integrated energy systems considering temporal dynamic and coupling characteristics.Energy, 195: 116964

[33] Chen B Y, Wang Y F (2021) Short-term electric load forecasting of integrated energy system considering nonlinear synergy between different loads.IEEE Access, 9: 43562-43573

[34] Vaghefi A, Jafari M A, Bisse E, et al.(2014) Modeling and forecasting of cooling and electricity load demand.Applied Energy, 136: 186-196

[35] Zhang L J, Shi J Q, Wang L L, et al.(2020) Electricity, heat,and gas load forecasting based on deep multitask learning in industrial-park integrated energy system.Entropy, 22(12): 1355

[36] Tan Z F, De G, Li M L, et al.(2020) Combined electricity-heatcooling-gas load forecasting model for integrated energy system based on multi-task learning and least square support vector machine.Journal of Cleaner Production, 248: 119252

[37] Zhou D J, Ma S X, Hao J R, et al.(2020) An electricity load forecasting model for Integrated Energy System based on BiGAN and transfer learning.Energy Reports, 6: 3446-3461

[38] Lindberg K B, Bakker S J, Sartori I (2019) Modelling electric and heat load profiles of non-residential buildings for use in longterm aggregate load forecasts.Utilities Policy, 58: 63-88

[39] Ge L J, Li Y L, Yan J, et al.(2021) Short-term load prediction of integrated energy system with wavelet neural network model based on improved particle swarm optimization and chaos optimization algorithm.Journal of Modern Power Systems and Clean Energy, 9(6): 1490-1499

Received: 23 October 2022/ Accepted: 28 February 2023/ Published: 25 June 2023

Fan Sun

15932266113@163.com

Yaojia Huo

Kinghuoyaojia@126.com

Huilan Liu

51312019@ncepu.edu.cn

Xi Wang

wx19991029@163.com

Lei Fu

leifuhappy@hbu.edu.cn

Yiming Ma

m13995218327@163.com

2096-5117/© 2023 Global Energy Interconnection Development and Cooperation Organization.Production and hosting by Elsevier B.V.on behalf of KeAi Communications Co., Ltd.This is an open access article under the CC BY-NC-ND license (http: //creativecommons.org/licenses/by-nc-nd/4.0/ ).

Biographies

Fan Sun, an assistant engineer, received master degree in electrical engineering from North China Electric Power University in 2021 and has been working in State Grid Ningxia Electric Power Technical Research Institute since 2021.His current research interest includes the load forecasting, coordinated operation of grid-power.

Yaojia Huo, an assistant engineer, received master degree in electrical engineering from China University of Mining & Technology in 2020 and has been working in State Grid Ningxia Electric Power Technical Research Institute since 2020.His current research interest includes electric power environmental protection technology and electric power technology innovation.

Lei Fu, a lecturer, received doctor degree from Yanshan University in 2019 and then has been working as a teacher in Hebei University.Her current research interest includes the analysis and control of time-delay singular systems,fuzzy systems, and synchronization control of complex networks.

Huilan Liu, a senior experimentalist, is currently studying for doctor degree in North China Electric Power University.Her current research interest includes equipment fault diagnosis and distributed energy storage technology research.

Xi Wang, a senior engineer, graduated from Shanghai Electric Power University and has worked in State Grid Ningxia Electric Power Technical Research Institute for more than 20 years.His current research interest includes automation technology.

Yiming Ma, a senior engineer, received master degree in electrical engineering from North China Electric Power University in 2014 and has been working in State Grid Ningxia Electric Power Technical Research Institute since 2014.His current research interest includes automation technology.

(Editor Dawei Wang)