A method and apparatus for predicting multiple loads in an integrated energy system

Abstract

This invention discloses a multi-dimensional load forecasting method for integrated energy systems, comprising the following steps: First, constructing a high-dimensional data matrix from the collected data using a split-window technique; Second, performing anomaly processing on outliers in the high-dimensional data matrix based on the Lonely Forest algorithm; Third, analyzing and repairing hidden anomalous data within the time series using the dynamic orbit method. If anomaly data is detected, the system triggers an alarm and enters the message surface behavioral analysis stage, visualizing the split window and identifying the anomaly type; Fourth, performing correlation analysis on multi-dimensional loads and their weather data using the MIC method to construct a high-dimensional feature matrix; Fifth, outputting the predicted values ​​of multi-dimensional loads based on a TLMMoE multi-task training network. The advantages of this invention are: This application provides a power load forecasting method and device, which can improve the accuracy of power load forecasting results.

Description

Technical Field

[0001] This invention relates to the technical field of integrated energy systems, and more specifically to the technical field of short-term multi-load forecasting for integrated energy systems. Background Technology

[0002] Environmental pollution and energy shortages have consistently constrained the sustainable development of human society. How to efficiently and economically achieve interconnectivity across different energy sectors, strengthen energy cooperation, and improve energy utilization has become a hot topic of widespread concern in the industry. To meet the urgent need for complementary and coordinated operation among various energy sources, Integrated Energy Systems (IES) have emerged. Short-term load forecasting is a prerequisite for IES to achieve intraday energy management and optimized scheduling; its forecasting accuracy directly affects the stability and economy of the IES during operation.

[0003] Load forecasting for integrated energy systems comprises four main stages: data feature preprocessing, model selection, optimization strategy, and algorithm selection. For a long time, the industry has conducted extensive and in-depth research on short-term load forecasting methods for integrated energy systems, achieving significant results. Existing forecasting methods mainly include traditional machine learning, ensemble learning, deep learning, reinforcement learning, and transfer learning. Addressing the problem of low prediction accuracy caused by the slow convergence speed of traditional wavelet neural network (WNN) integrated energy system load forecasting models, a short-term load forecasting method for integrated energy systems based on improved particle swarm optimization (IPSO) is proposed. Considering the coupling relationship between heat, electricity, and gas loads in industry, this paper analyzes multi-dimensional data such as weather information and historical data to mine the coupling relationships between data, proposing a prediction method based on deep learning and multi-task learning to improve the model's predictive ability and obtain more accurate prediction results. Considering the high-dimensional temporal dynamics, an encoder-decoder model based on long short-term memory networks (LSTMED) is proposed, verifying that deep learning methods can be well applied to integrated energy load forecasting. By analyzing the coupling relationships between different subsystems, and utilizing the concepts of multi-task learning weight sharing mechanism and least squares support vector machine, a joint forecasting method for electricity, heat, cooling, and gas loads based on multi-task learning and least squares support vector machine is proposed. A comprehensive load forecasting model based on bidirectional generative adversarial network (Bi-GAN) data augmentation and transfer learning techniques is also proposed. This model considers the data scarcity in the information system, maintains high forecasting accuracy, addresses the issue of new users joining the integrated energy system, improves the robustness of the integrated energy system, and expands the generalization of the integrated energy system.

[0004] In recent years, with the large-scale integration of various intelligent systems and distributed energy sources, energy interaction structures have become increasingly diversified, and user energy consumption characteristics under these multi-structures have become more real-time and complex. In actual operation, integrated energy systems may operate under extreme conditions such as sudden changes in the external environment and equipment failures, resulting in load data that is considered unconventional. Due to the limited amount of data under unconventional modes, parameter optimization is difficult, and neural networks are highly sensitive to abnormal data. When the load is affected by complex factors and exhibits strong randomness and non-stationarity, the prediction accuracy of neural networks will significantly decrease. To date, most researchers have focused on conventional design for integrated energy systems in the model selection, optimization strategy, and algorithm selection stages, which cannot meet the operational needs of integrated energy systems with diverse modes and sudden changes in operating conditions. Therefore, on the one hand, design should be implemented in the data feature preprocessing stage to remove abnormal data. Simultaneously, to meet the needs of actual operating conditions, the impact of event-driven factors needs to be considered, and the historical energy consumption characteristics of users need to be analyzed to improve the overall stability of the system. On the other hand, appropriate parameters should be selected in the model selection, optimization strategy, and algorithm selection stages. Through multi-stage scheme design, the load prediction accuracy under conventional modes can be improved.

[0005] Currently, methods for outlier identification in time-series data are constantly evolving. Common cleaning methods include statistical methods such as Gaussian distribution, box plots, and clustering methods from machine learning. One method, based on whether the measurement data error follows a zero-mean Gaussian distribution, uses the "3σ criterion" to detect outliers. While this method is relatively simple and convenient, it suffers from significant limitations in the operation of integrated energy systems, which are influenced by various factors, resulting in high-dimensionality, randomness, intermittency, and uncertainty in the output data. The accuracy of this method for identifying outliers is greatly reduced, only able to identify simple outliers. Clustering algorithms, as a type of machine learning algorithm, can identify most outliers, but they struggle to identify hidden outliers within the sequence, failing to meet the requirements for high-quality data analysis. Summary of the Invention

[0006] In view of this, the present invention proposes a method and apparatus for predicting multiple loads in an integrated energy system, which can eliminate abnormal data in a time series under a conventional data mode and is applicable to the prediction of multiple loads in an integrated energy system.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0008] A multi-element load forecasting method for integrated energy systems includes the following steps:

[0009] The first step is to construct a high-dimensional data matrix from the collected data using the split window technique;

[0010] The second step is to perform anomaly processing on outliers in the high-dimensional data matrix based on the Lonely Forest algorithm.

[0011] The third step is to analyze and repair the abnormal data hidden in the time series using the dynamic track method. If abnormal data is found, the system will trigger an alarm and enter the message surface behavioral analysis stage, and use the visualization separation window to identify the type of abnormal value.

[0012] The fourth step is to conduct correlation analysis on multivariate loads and their weather data using the MIC method to construct a high-dimensional feature matrix;

[0013] The fifth step is to train the network based on TLMMoE to output predicted values ​​of multi-variable loads.

[0014] Preferably, in step one, a high-dimensional data matrix is ​​constructed from the collected data using a split-window technique, based on historical cooling load, heating load, electrical load, and meteorological data.

[0015] (1) Construction of high-dimensional data matrix: The first step in data feature preprocessing is to construct a reasonable high-dimensional data matrix X. T First, construct N-dimensional nodes for data sampling, selecting cold, heat, electricity load, and weather data. At a certain sampling time t, the N-dimensional sampling nodes can form a column vector:

[0016] x t =[ x 1,t x 2,t … x N,t ] T (1)

[0017] As the sampling time increases, the column vector data also increases, thus forming the following time series X:

[0018] X = [x1 x2 … x t …] (2)

[0019] To meet the needs of real-time computing, the collected data is sequentially extracted and analyzed using a sliding time window. The width of the sliding time window is set to τ, which means collecting data at time t and historical data τ-t before time t. A high-dimensional data matrix X is constructed from all the data within this time window. T as follows:

[0020] X T =[ x t-τ+1 … x t (3)

[0021] In the formula, τ is the time window.

[0022] The selection of the time window should meet the following conditions: Set a time window of appropriate width so as to meet the needs of real-time detection and also include all measurement data of all partitions;

[0023] (2) Data standardization: Constructing a high-dimensional data matrix X T All data is standardized, that is, dedimensionalized and normalized.

[0024] Preferably, in step two, outlier handling is performed on the high-dimensional data matrix based on the Lonely Forest algorithm:

[0025] By constructing a random forest, the isolation level of data points is estimated to determine whether they are outliers. The Lonely Forest algorithm is particularly sensitive to outliers in high-dimensional data matrices composed of cooling load, heating load, electrical load, and meteorological data in integrated energy systems, and can efficiently and quickly isolate and separate them.

[0026] Typically, the basic steps of the Lonely Forest algorithm are as follows:

[0027] The first step is to randomly select one feature and one cut point within the range of feature values;

[0028] The second step is to use the selected features and cut points as segmentation rules to divide the data points into two subsets.

[0029] The third step is to recursively repeat the first and second steps until there is only one data point in each subset, or the maximum depth of the tree is reached.

[0030] The fourth step is to construct multiple such random trees and form a random forest;

[0031] Fifth step: For each data point, calculate its path length in the random forest, which is the average number of edges traversed from the root node to the data point.

[0032] The sixth step is to use path length to measure the degree of anomaly of data points. Shorter path lengths indicate normal points that are easier to segment, while longer path lengths indicate anomalies that are easier to isolate.

[0033] The seventh step involves setting a threshold to compare the path length with the probability of outliers, thereby determining which data points are classified as outliers.

[0034] Preferably, in step three, the abnormal data hidden in the time series is analyzed and repaired using the dynamic trajectory method. If abnormal data is found, the system triggers an alarm and enters the message surface behavioral analysis stage. A visualization separation window is used to determine the type of abnormal value. The specific steps are as follows:

[0035] (1) This invention proposes a Dynamic Railway Line (DRL) method applicable to integrated energy systems to repair data anomalies. The dynamic railway line consists of a base rail (Middle Railway Line, MRL), an upper rail (Upper Railway Line, URL), and a lower rail (Lower Railway Line, LRL). The shape of the dynamic railway line looks like a channel. When the data value is within the channel range, it is considered a normal value; when the data value crosses the channel and appears outside the channel, it is determined to be an anomaly.

[0036] (2) The calculation of the dynamic orbit method mainly consists of two steps:

[0037] 1) First, the base orbit of the dynamic orbit is established using the moving average method;

[0038] 2) Next, the upper and lower rails are calculated. Based on the determined base rail, the upper and lower rails are calculated using the following formulas:

[0039] URL = (1 + M1) * MRL (4)

[0040] LRL=(1-M2)*MRL (5)

[0041] In the formula, M1 is the margin value of the upper track, and M2 is the margin value of the lower track;

[0042] Preferably, the present invention selects a weighted moving average indicator. Starting with the data from the first time point in the time series, the weight assigned to the time series data increases linearly as the time series grows.

[0043] The dynamic orbit method is used to remove and repair outliers in the time series. When an outlier exceeds the upper orbit value of the dynamic orbit, the base orbit value is assigned to the corresponding value of the high-dimensional data matrix; when an outlier exceeds the lower orbit value of the dynamic orbit, the base orbit value is assigned to the corresponding value of the high-dimensional data matrix.

[0044] (3) When an abnormal alarm occurs, the system’s message surface behavioral analysis function will be triggered. The system will analyze the activity status within the system to determine whether it is due to a sudden event, such as a large-scale event or a new building and its equipment being put into use. It will also provide experience values ​​of extreme events for future production and life, and help develop integrated prediction algorithms.

[0045] Preferably, in step four, correlation analysis is performed on multivariate loads and their weather data using the MIC method to construct a high-dimensional feature matrix:

[0046] (1) Correlation analysis of multiple loads:

[0047] An integrated energy system is a complex system composed of cooling, heating, and electrical systems as its core, and combined with meteorological and economic systems. As a complex, large-scale, and dynamic time-varying system, various energy sources within the system are coupled and complement each other. At the same time, with the continuous expansion of the scale and increasing complexity of the integrated energy system, the time dimension of the integrated energy system operation data has increased, resulting in massive amounts of data with time-series state characteristics and spatiotemporal dimension characteristics. Analyzing the collected high-dimensional characteristic data helps to gain a general understanding of the entire integrated energy system and qualitatively points out the direction and scope for the construction of prediction models.

[0048] (2) Autoregressive system analysis:

[0049] This invention employs an autoregressive model (AR model) to analyze the temporal state characteristics of short-term time series, providing a basis for the moving average number of days l of the base orbit in the dynamic orbit method. The autoregressive model uses itself as the regression variable, that is, it uses a linear combination of its own information data from previous periods to describe the linear regression relationship of a random variable at a future point in time. A time series {Xt} is defined, satisfying:

[0050] X t =a0+aX t-1 +…+a p X t-p +ε t (6)

[0051] In the formula, a0, a1, ..., a p It is the coefficient, ε t The values ​​are white noise. The time-series characteristics of the load data are analyzed based on the autocorrelation coefficient (ACF) and partial autocorrelation coefficient (PACF) of the AR model.

[0052] (3) Maximum Information Coefficient Analysis

[0053] The MIC method can measure the strength of linear and nonlinear correlations between data. By evaluating the correlations between different types of load data, it discovers a wide range of relationship types. Therefore, this invention uses the Maximum Information Coefficient (MIC) to perform spatiotemporal characteristic analysis on multiple systems to construct a high-dimensional feature matrix for the model input. The correlation between the major features ranges from [0,1]. A value less than 0.3 indicates a weak correlation, while a value greater than 0.7 indicates a strong correlation.

[0054] (4) Construction of input feature set

[0055] Based on MIC (Multidimensional Correlation Analysis), multidimensional correlation analysis is performed on high-dimensional data. The following data features are used as input data, with a granularity of 15 min:

[0056] Input1: Predicted cooling load one year prior to the base date;

[0057] Input2: Predicted heat load one year before the base date;

[0058] Input3: Predicted electrical load one year prior to the base date;

[0059] Input4: Predict temperature data for one year prior to the baseline date;

[0060] Input5: Dew point data predicted one year prior to the base date;

[0061] Input6: Predicted building data for cooling one year prior to the base date;

[0062] Input7: Predicted building heat consumption data one year prior to the base date;

[0063] Input8: Predict building electricity consumption data one year prior to the base date.

[0064] Preferably, in step five, the predicted values ​​of the multi-variable workloads are output based on the TLMMoE multi-task training network:

[0065] This invention employs the TCN-LSTM-MMoE multi-task learning framework, hereinafter referred to as the TLMMoE model, to establish a multivariate load forecasting model.

[0066] (1) When the network model receives the input high-dimensional feature matrix, it uses a temporal convolutional neural network (TCN) to mine and extract the temporal and spatiotemporal characteristics of the high-dimensional data features;

[0067] (2) The TCN network inputs the learned feature information into the Long Short-Term Memory (LSTM) network to further learn the feature matrix and uses the unique memory function of the LSTM unit to extract the temporal links of the feature matrix.

[0068] (3) The processed data features are input into the MMoE (Multi-gate Mixture-of-Experts) multi-task learning model, and the power values ​​of cooling load, heating load and electrical load are predicted and output using the expert network layer and unique gating mechanism of the MMoE model.

[0069] This invention provides a multi-dimensional load forecasting device for integrated energy systems. The device includes: a high-dimensional data receiving module, a high-dimensional data preprocessing module, a multi-dimensional feature analysis module, a data feature mining module, a multi-dimensional feature refinement and extraction module, and a data information prediction output module.

[0070] The high-dimensional data receiving module collects data sets from the measurement device and builds a high-dimensional data matrix based on the split window technology, with the split window sliding at intervals.

[0071] The high-dimensional data preprocessing module uses the Lonely Forest algorithm and the Dynamic Track method to remove outliers and abnormal values ​​in the time series.

[0072] The multi-dimensional feature analysis module uses the MIC analysis method to perform correlation analysis on the cooling load, heating load, electrical load and meteorological data of the dataset, and select input features;

[0073] The data feature mining module uses the temporal convolutional unit in the TCN network layer to mine and extract temporal and spatial features from the time-series data.

[0074] The multi-dimensional feature refinement extraction module uses the memory units in the LSTM network layer to refine the input data of the previous layer and mine the long short-term memory relationship.

[0075] The data information prediction output module uses the expert network layer of the MMoE model and a unique gating mechanism to predict and output the power values ​​of cooling load, heating load, and electrical load.

[0076] The present invention also provides a computer device for a multi-load forecasting device for an integrated energy system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the power load forecasting method described above.

[0077] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the power load forecasting method described above.

[0078] The beneficial effects that can be achieved by adopting the above-mentioned technical solution in this invention are:

[0079] This application provides a method and apparatus for predicting power load, which can improve the accuracy of power load prediction results. Addressing the problem of multi-element load prediction in integrated energy systems, this application designs the data feature preprocessing and model optimization stages, proposing a multi-element load prediction method for integrated energy systems based on the Lonely Forest algorithm and the Dynamic Track method.

[0080] Based on the analysis of examples, the following conclusions were drawn:

[0081] 1) The Lonely Forest algorithm can handle outlier problems in high-dimensional big data, and has the characteristics of high processing accuracy and fast calculation speed;

[0082] 2) The dynamic trajectory method can effectively remove hidden outliers in time series, with good cleaning effect, providing a good foundation for neural network prediction models;

[0083] 3) The coupling relationship between load data in integrated energy systems is complex. The maximum information coefficient analysis method can better analyze the temporal and spatial characteristics between loads. By analyzing and constructing a highly correlated high-dimensional feature matrix, it lays the foundation for improving the accuracy of the prediction model.

[0084] 4) Through the rational design of the TLMMoE network structure, starting from the three stages of data feature capture, learning and multi-task allocation, the algorithm learns the coupling characteristics of historical data better, improves the prediction accuracy, and proves the effectiveness of the algorithm in time-series feature sequences.

[0085] The findings of this study provide a new approach and method for predicting multi-variable loads in integrated energy systems, which is of great significance for promoting the intelligentization and efficiency of integrated energy systems. Attached Figure Description

[0086] Figure 1 This is a diagram showing the electrical load operation before the processing described in this invention;

[0087] Figure 2 This is a contour map showing the electrical load distribution of the present invention.

[0088] Figure 3 This is a diagram showing the electrical load operation after processing according to the present invention;

[0089] Figure 4 This is an ACF and PACF analysis diagram of the cooling load of this invention;

[0090] Figure 5 This is an ACF and PACF analysis diagram of the electrical load of the present invention;

[0091] Figure 6 This is a heatmap of the characteristic sequence MIC of the present invention;

[0092] Figure 7 This is a model framework diagram of the present invention;

[0093] Figure 8 This is a comparison chart of the errors between the IDTLMMoE model and the ITLMMoE model of this invention;

[0094] Figure 9 This is a comparison chart of the cooling loads of the IDCLSTM model and the IDTLMMoE model of this invention.

[0095] Figure 10 This is a comparison chart of the heat load of the IDCLSTM model and the IDTLMMoE model of this invention;

[0096] Figure 11 This is a comparison diagram of the electrical load of the IDCLSTM model and the IDTLMMoE model of this invention;

[0097] Figure 12 This is a comparison chart of the errors between the IDTLMMoE model and the IDCLSTM model of this invention. Detailed Implementation

[0098] The load data for the simulation example comes from the IES cooling, heating, and electricity load data and weather characteristics of Arizona State University's Tempe campus from January 1, 2022 to May 1, 2023, with a sampling granularity of 15 minutes. The load data is sourced from the campus's project network database. The campus has 288 buildings, over 50,000 faculty and students, and possesses combined cooling, heating, and power (CCHP), electric boilers, gas boilers, and power-to-gas (P2G) energy conversion equipment. Meteorological data is sourced from publicly available meteorological data for the corresponding time period on meteorological websites. The dataset is divided as follows: 75% for training, 15% for validation, and 10% for testing.

[0099] The experimental platform was a hardware platform with a Xeon 4213R CPU and an NVIDIA RTX 3090 GPU. The TensorFlow and Keras deep learning frameworks were built and implemented using the Python language.

[0100] 1. Construction of high-dimensional data matrices based on split window technique

[0101] A high-dimensional data matrix is ​​constructed based on historical cooling load, heating load, electrical load, and meteorological data.

[0102] Construction of a high-dimensional data matrix: The first step in data feature preprocessing is to construct a reasonable high-dimensional data matrix Xp. In this invention, a 21-dimensional node is first constructed to sample cold, heat, electricity load, and weather data. Data feature information is detailed below. Figure 6 .

[0103] At a certain sampling time t1, the 21-dimensional sampling nodes can form a column vector:

[0104] x t =[ x 1,t x 2,t … x 21,t1 ] T (1)

[0105] As the sampling time increases, the column vector data also increases, thus forming the following time series X:

[0106] X = [x1 x2 … x 21 (2)

[0107] To meet the needs of real-time computing, the collected data is sequentially extracted and analyzed using a sliding time window. The width of the sliding time window is set to τ, which includes the data collected at time t1 and the historical data τ-t1 before time t1. A high-dimensional data matrix X is constructed from all the data within this time window. T as follows:

[0108] X T =[ x t1-τ+1 … x t1 (3)

[0109] In the formula, τ is the time window.

[0110] The selection of the time window should meet the following conditions: Set a time window of appropriate width so as to meet the needs of real-time detection and also include all measurement data of all partitions.

[0111] (2) Data standardization: Standardize all data, that is, remove dimensions and normalize.

[0112] The data in the examples all have different dimensions. For example, the unit for cooling load is Ton / hrs, with a value ranging from 1000 to 9000; the unit for heating load is mmBTU / hr, with a value ranging from 1 to 30; and the unit for electrical load is kW, with a value ranging from 7000 to 26000. It can be seen that the three different loads have different dimensional values, which is detrimental to the convergence of the neural network load prediction model. Therefore, all data are standardized. After normalization, the processed data for each feature conforms to a standard normal distribution, i.e., the mean is 0 and the standard deviation is 1.

[0113] 2. Outlier Identification Based on Lonely Forest Algorithm

[0114] The example data selected in this invention includes 21 statistical features and over 46,000 data points. Faced with this massive amount of experimental data, this invention uses the Lonely Forest algorithm to process the global data. This algorithm is characterized by high processing accuracy and fast computation speed, making it suitable for handling outlier problems in large datasets. Furthermore, the processed data is more conducive to the implementation of the dynamic orbit method.

[0115] Due to space limitations, this paper only uses electrical load data as an example to demonstrate the processing procedure and provides a necessary analysis. Figure 1 The image shows a historical data chart of the electrical load.

[0116] Observations revealed extreme values ​​in the data, causing the dynamically changing power load data to appear as a straight line, and there was a difference in magnitude between the extreme values ​​and the normal data.

[0117] To more clearly illustrate this magnitude relationship, a contour map of the electrical load was drawn, such as... Figure 2 As shown, it can be clearly observed that the white dots in the image represent normal values, distributed within the contour line clusters in the blue area. The closer to the center, the darker the color; the further outward from the contour lines, the lighter the color. The black dots in the image represent outliers, scattered within the dark green area.

[0118] By detecting and removing outliers, the actual operating data of the power load after removal is obtained, such as... Figure 3 As shown in the figure, by comparing the operation charts before and after power load preprocessing, it can be seen that the outliers in the data have been removed, and the power load operation data shows a trend and regularity, indicating that the method can efficiently handle outlier points in the data.

[0119] 3. Analyze and repair abnormal data hidden within the time series using the dynamic trajectory method. If abnormal data is detected, the system triggers an alarm and enters the message surface behavioral analysis stage, where a visual separation window is used to identify the type of abnormal value.

[0120] (1) First, the parameter l is determined by autoregression of the data using the AR algorithm, thereby determining the base orbit value of the dynamic orbit.

[0121] Due to space limitations, this article will use electrical load data and cooling load data as examples. Figure 4 The diagram shows the autocorrelation coefficient analysis and partial autocorrelation coefficient analysis of cooling load, as follows: Figure 5 The diagram shows the autocorrelation coefficient analysis and partial autocorrelation coefficient analysis of electrical load.

[0122] like Figure 4 As shown, the blue area represents the confidence interval of the correlation coefficient plot. In this paper, a 95% confidence interval is used, with a time granularity of 15 minutes. The autocorrelation coefficient plot of cooling load shows a strong coupling correlation between loads within the first 13 time intervals, considering all influencing factors. The partial autocorrelation plot of cooling load shows a strong coupling correlation between loads within the first 3 time intervals. Comparing the autocorrelation coefficient plots of cooling load and electrical load reveals that cooling load exhibits a higher short-term trend than electrical load. Furthermore, both cooling load and electrical load demonstrate periodicity. Therefore, the l-value for cooling load is set to 8, and the l-value for electrical load is set to 3.

[0123] (2) Next, calculate the upper and lower rails. The specific calculation formula is as follows:

[0124] URL = (1 + M1) * MRL (4)

[0125] LRL=(1-M2)*MRL (5)

[0126] In the formula, M1 is the margin value of the upper track (10%), and M2 is the margin value of the lower track (10%).

[0127] Preferably, the present invention selects a weighted moving average indicator. Starting with the data from the first time point in the time series, the weight assigned to the time series data increases linearly as the time series grows.

[0128] The dynamic orbit method is used to remove and repair outliers in the time series. When an outlier exceeds the upper orbit value of the dynamic orbit, the base orbit value is assigned to the corresponding value of the high-dimensional data matrix; when an outlier exceeds the lower orbit value of the dynamic orbit, the base orbit value is assigned to the corresponding value of the high-dimensional data matrix.

[0129] (3) When an abnormal alarm occurs, the system’s message surface behavioral analysis function will be triggered. The system will analyze the activity status within the system to determine whether it is due to a sudden event, such as a large-scale event or a new building and its equipment being put into use. It will also provide experience values ​​of extreme events for future production and life, and help develop integrated prediction algorithms.

[0130] 4. Correlation analysis of multivariate loads and their weather data is performed using the MIC method to construct a high-dimensional feature matrix:

[0131] (1) Correlation analysis of multiple loads:

[0132] This embodiment uses MIC analysis to analyze and describe the correlation between various variables.

[0133] Pearson correlation coefficient or Spearman correlation coefficient can be used to effectively measure the linear correlation of data, and even mathematical formulas for linear and simple nonlinear relationships can be determined through regression analysis. However, in integrated energy systems, the coupling relationships between multiple systems are complex, and there are not only linear relationships between data, but also a large number of nonlinear relationships, which cannot be simply expressed by mathematical formulas.

[0134] The MIC method can measure the strength of linear and nonlinear correlations between data. By evaluating the correlations between different types of load data, it discovers a wide range of relationship types. Therefore, this invention uses the Maximum Information Coefficient (MIC) to perform spatiotemporal characteristic analysis on multiple systems to construct a high-dimensional feature matrix for the model input. Figure 6 The diagram shows the MIC heat map between multiple systems, including the power system, cooling system, heating system, and weather system.

[0135] Figure 6 The legend on the right side of the figure shows the correlation between the major features, with values ​​ranging from [0,1]. When the value is less than 0.3, it indicates a weak correlation between the two features; when the value is greater than 0.7, it indicates a strong correlation between the two features.

[0136] As shown in the figure, the cooling system, heating system, and electrical system are the three basic systems in the integrated energy system. The correlation coefficient between cooling load and heating load is 0.57, indicating a strong correlation and a close coupling relationship between them. The correlation coefficient between cooling load and electrical load is 0.7, demonstrating a particularly close connection between the two systems in production and daily life within the integrated energy system.

[0137] The correlation coefficient between the thermal and electrical systems is 0.48, indicating that a coupling relationship exists between them in the integrated energy system, although it is relatively weaker than that between the cooling and electrical systems. Furthermore, it can be observed that the cooling, heating, and electrical systems are all closely correlated with temperature and dew point. In addition, the building usage of each system shows a strong correlation with its respective system.

[0138] (2) Construction of input feature set

[0139] Based on MIC (Multidimensional Correlation Analysis), multidimensional correlation analysis is performed on high-dimensional data. Considering factors such as model optimization speed, the following data features are used as input data (granularity is 15 min):

[0140] Input1: Predicted cooling load one year prior to the base date;

[0141] Input2: Predicted heat load one year before the base date;

[0142] Input3: Predicted electrical load one year prior to the base date;

[0143] Input4: Predict temperature data for one year prior to the baseline date;

[0144] Input5: Dew point data predicted one year prior to the base date;

[0145] Input6: Predicted building data for cooling one year prior to the base date;

[0146] Input7: Predicted building heat consumption data one year prior to the base date;

[0147] Input8: Predict building electricity consumption data one year prior to the base date.

[0148] The detailed list is shown in Table 1:

[0149]

[0150] 5. Based on TLMMoE multi-task training network, output the predicted values ​​of multivariate loads.

[0151] This invention employs the TCN-LSTM-MMoE multi-task learning framework, hereinafter referred to as the CLMMOE model, to establish a multivariate load forecasting model. The model framework is as follows: Figure 7 As shown, firstly, after the network model receives the input high-dimensional feature matrix, the TCN network mines and extracts the temporal and spatiotemporal characteristics of the high-dimensional data features. Next, the TCN network inputs the learned feature information into the LSTM network for further learning of the feature matrix, utilizing the unique memory function of the LSTM units to extract temporal links from the feature matrix. Then, the processed data features are fed into the MMoE multi-task learning model, where the expert network layer and unique gating mechanism of the MMoE model are used to predict and output the power values ​​of cooling load, heating load, and electrical load.

[0152] (1) Evaluation indicators

[0153] This invention uses Mean Absolute Percentage Error (MAPE) as the primary evaluation index for the predictive performance of each model, and selects Mean Absolute Error (MAE) as an auxiliary evaluation index. The calculation formulas for MAPE and MAE are as follows:

[0154]

[0155]

[0156] In the formula: y t This is actual power load data; Forecasted power load data; is the mean of the power load; n is the number of samples of the power load data.

[0157] (2) Hyperparameter settings

[0158] Hyperparameters exist in the TCN network layer, LSTM network layer, and multi-task learning layer of the network model. Based on the learning effect of model training, the hyperparameters in the model are determined using the control variable method. After multiple parameter adjustments, the final optimized parameters are shown in Table 2.

[0159]

[0160] 6. Performance analysis of different prediction models

[0161] The method proposed in this invention differs from traditional multi-task methods in both data preprocessing and model structure. To further comprehensively evaluate the correctness and effectiveness of the proposed method based on the Lonely Forest algorithm and dynamic orbit method in multi-variable load forecasting of integrated energy systems, this section objectively designs three sets of experimental models:

[0162] 1) Based on the iForest-TCN-LSTM-MMoE multivariate load forecasting model (ITLMMoE model);

[0163] 2) Based on the iForest-DRL-CNN-LSTM multivariate load forecasting model (IDCLSTM model);

[0164] 3) Based on the iForest-DRL-TCN-LSTM-MMoE multivariate load forecasting model (IDTLMMoE model).

[0165] To ensure fairness in the experiment, all models used the same dataset allocation ratio to divide the training, validation, and test sets. Furthermore, to guarantee the training performance of the comparison models, the network layers were configured optimally.

[0166] (1) Comparison of IDTLMMoE model and ITLMMoE model

[0167] To verify that the IDTLMMoE model proposed in this invention can remove outliers during data preprocessing and improve load forecasting accuracy, a control model, the IDTLMMoE model, is set up. The control model uses only the Lonely Forest algorithm for data preprocessing, and then uses the TLMMoE network layer for load forecasting.

[0168] like Figure 8 The image shows the errors of the two models on the test set. It is evident that adding the DRL module significantly reduces the prediction errors for both cooling and electrical loads. This is because the Dynamic Tracking (DRL) method effectively removes outliers hidden in the time series, thereby reducing the interference of outlier points on the neural network and improving prediction accuracy.

[0169] Table 3 shows a comparison of the prediction errors of the IDTLMMoE model and the ITLMMoE model proposed in this invention. The PAPE indices for cooling load, heating load, and electrical load decreased by 2.57%, 0.11%, and 0.49%, respectively, and the MAE error values ​​for cooling load, heating load, and electrical load also decreased to some extent. Therefore, the model proposed in this invention is better, confirming that the dynamic trajectory method can indeed remove hidden outliers in time series, reduce prediction errors, and improve load prediction accuracy.

[0170]

[0171] (2) Comparison of IDTLMMoE and ITLLMMoE models

[0172] To verify the effectiveness of the proposed IDTLMMoE model in the TLMMoE module of the network layer structure, this invention sets up a control model, the IDCLSTM model. The control model is a traditional multi-task load prediction model based on hard parameter sharing. The traditional multi-task learning mechanism uses an LSTM layer as the sharing layer, and then connects it to a fully connected layer as the prediction output, thereby achieving multi-variable load prediction.

[0173] like Figure 9 , Figure 10 and Figure 11As shown, we present a performance comparison of the IDCLSTM and IDTLMMoE models on a test set for a selected period of time for cooling, heating, and electrical loads. It can be seen that in the electrical load prediction, both the proposed IDTLMMoE model and the traditional multi-task IDCLSTM model achieve good prediction results, with the predicted curves fitting the actual curves well. The IDTLMMoE model shows better prediction performance than the IDCLSTM model. However, through a detailed comparison, the IDTLMMoE model maintains a good fit in the parts of the curve with significant fluctuations, and the predicted values ​​are closer to the actual values, while the IDCLSTM model's prediction error increases, and its fit is poor.

[0174] Furthermore, because the fluctuations in cooling and heating loads are more drastic than those in electrical loads, model training is more difficult, and overfitting or underfitting is more likely to occur. Figure 12 As can be seen, the IDCLSTM model is more conservative in its predictions of cooling and heating loads, and its prediction curves exhibit smoother fluctuations when faced with highly volatile cooling and heating loads. The MAPE percentage for cooling load in the IDCLSTM model is 6.98%, while the IDTLMMoE model maintains good predictive ability with a MAPE percentage of 3.8%, a reduction of 3.18 percentage points compared to the IDCLSTM model. This is because using LSTM as a hard-shared layer cannot account for the correlation differences between multiple loads, leading to the provision of the same weight values ​​to the sub-task layers, thus affecting model performance. In contrast, the proposed IDTLMMoE model achieves feature information sharing among multiple tasks by setting up multiple expert subnets. Simultaneously, the gating layer avoids interference from weakly correlated information on sub-tasks. Furthermore, the IDTLMMoE model sets different loss functions based on the volatility characteristics of sub-tasks and the correlation differences between multiple loads. Therefore, the IDTLMMoE model achieves better prediction results.

[0175] like Figure 12 As can be seen, the IDTLMMoE model proposed in this invention outperforms the IDCLSTM model in terms of performance on the test set for cold, heat, and electrical loads, with reduced MAPE error percentage and MAE error value. This also indicates that the TCN layer in the IDTLMMoE model is better able to acquire long-term sequence feature information and perform data mining processing compared to the CNN layer in the IDCLSTM model. Consequently, when the feature information is fed into the LSTM layer, the LSTM unit layer in the IDTLMMoE model is better able to refine and learn these temporal and spatial features. Therefore, the IDTLMMoE model has better prediction performance.

Claims

1. A multi-element load forecasting method for integrated energy systems, characterized in that, Includes the following steps: The first step is to construct a high-dimensional data matrix from the collected data using the split window technique; The second step is to perform anomaly processing on outliers in the high-dimensional data matrix based on the Lonely Forest algorithm. The third step is to analyze and repair the abnormal data hidden in the time series using the dynamic track method. If abnormal data is found, the system will trigger an alarm and enter the message surface behavioral analysis stage, and use the visualization separation window to identify the type of abnormal value. (1) A dynamic track method DRL suitable for integrated energy systems is proposed to repair data anomalies. The dynamic track consists of a base track MRL, an upper track URL, and a lower track LRL. The shape of the dynamic track looks like a channel. When the data value is within the channel range, it is considered a normal value; when the data value crosses the channel and appears outside the channel, it is judged as an anomaly. (2) The calculation of the dynamic orbit method mainly consists of two steps: 1) First, the base orbit of the dynamic orbit is established using the moving average method; 2) Next, the upper and lower rails are calculated. Based on the determined base rail, the upper and lower rails are calculated using the following formulas: (4) (5) In the formula, M1 is the margin value of the upper track, and M2 is the margin value of the lower track; The dynamic orbit method is used to remove and repair outliers in the time series. When an outlier exceeds the upper orbit value of the dynamic orbit, the base orbit value is assigned to the corresponding value of the high-dimensional data matrix; when an outlier exceeds the lower orbit value of the dynamic orbit, the base orbit value is assigned to the corresponding value of the high-dimensional data matrix. The fourth step is to conduct correlation analysis on multivariate loads and their weather data using the MIC method to construct a high-dimensional feature matrix; Step 5, based on The multi-task training network outputs predicted values ​​for diverse workloads. based on The multi-task training network outputs predicted values ​​for various workloads: use Multi-task learning framework, hereinafter referred to as Model, establish a multivariate load forecasting model, (1) After the network model receives the input high-dimensional feature matrix, it uses the TCN network to mine and extract the temporal and spatiotemporal characteristics of the high-dimensional data features; (2) The TCN network inputs the learned feature information into the LSTM network to further learn the feature matrix and uses the unique memory function of the LSTM unit to extract the temporal links of the feature matrix. (3) Input the processed data features into In multi-task learning models, utilizing The model's expert network layer and unique gating mechanism predict and output the power values ​​of cooling load, heating load, and electrical load.

2. The method for predicting multiple loads in an integrated energy system according to claim 1, characterized in that, In the first step, a high-dimensional data matrix is ​​constructed from the collected data using the split window technique. This high-dimensional data matrix is ​​built based on historical cooling load, heating load, electrical load, and meteorological data. (1) Construction of high-dimensional data matrix: First, construct an R-dimensional data matrix by selecting cold, heat, electricity load and weather data nodes for data sampling. At a certain sampling time R, the R-dimensional sampling nodes can form a column vector: (1) As the sampling time increases, the column vector data also increases, thus forming the following time series 𝑋: (2) The collected data is sequentially extracted and analyzed using a sliding time window. The width of the sliding time window is set to 𝜏, meaning it is used to capture data at time 𝑡 and before time 𝑡. Historical data, using all data within this time window to construct a high-dimensional data matrix. as follows: (3) In the formula, 𝜏 represents the time window; (2) Data standardization: Constructing a high-dimensional data matrix All data is standardized, that is, dedimensionalized and normalized.

3. The method for predicting multiple loads in an integrated energy system according to claim 1, characterized in that, In the second step, outlier values ​​in the high-dimensional data matrix are handled using the Lonely Forest algorithm: By constructing a random forest, the isolation level of data points is estimated, thereby determining whether they are outliers. The Lonely Forest algorithm is highly sensitive to outliers in the high-dimensional data matrix of integrated energy systems, which consists of cooling load, heating load, electrical load, and meteorological data. It can efficiently and quickly isolate and separate them.

4. The method for predicting multiple loads in an integrated energy system according to claim 1, characterized in that, In the third step, the dynamic trajectory method is used to analyze and repair abnormal data hidden in the time series. If abnormal data is found, the system triggers an alarm and enters the message surface behavioral analysis stage, where a visualization separation window is used to identify the type of abnormal value. When an abnormal alarm occurs, the system's message-side behavioral analysis function will be triggered. The system will analyze the activity status within the system to determine whether it is due to a sudden event, and provide experience values ​​for extreme events for future production and life, and help develop integrated prediction algorithms.

5. The method for predicting multiple loads in an integrated energy system according to claim 1, characterized in that, In the fourth step, correlation analysis is performed on multivariate loads and their weather data using the MIC method to construct a high-dimensional feature matrix: (1) Correlation analysis of multiple loads: An integrated energy system is a complex system composed of cooling, heating, and electrical systems as its core, and combined with meteorological and economic systems. As a complex, large-scale, and dynamic time-varying system, various energy sources within the system are coupled and complement each other. At the same time, with the continuous expansion of the scale and increasing complexity of the integrated energy system, the time dimension of the integrated energy system operation data has increased, resulting in massive amounts of data with time-series state characteristics and spatiotemporal dimension characteristics. Analyzing the collected high-dimensional characteristic data helps to gain a general understanding of the entire integrated energy system and qualitatively points out the direction and scope for the construction of prediction models. (2) Autoregressive system analysis: An autoregressive (AR) model is used to analyze the time-series state characteristics of short-term time series data, providing a basis for the moving average number of days (l) of the base orbit in the dynamic orbit method. The AR model uses itself as the regression variable, that is, it uses a linear combination of its own information data from previous periods to describe the linear regression relationship of a random variable at a future point in time. A time series is defined... ,satisfy: (6) In the formula , ,⋯, It is a coefficient. The values ​​are white noise. The time-series state characteristics of the load data are analyzed based on the autocorrelation function ACF and the partial autocorrelation function PACF of the AR model. (3) Analysis of maximum information coefficient The MIC method can measure the strength of linear and nonlinear correlations between data. By evaluating the correlation between different types of load data, a wide range of relationship types can be discovered. Therefore, the maximum information coefficient (MIC) is used to perform spatiotemporal dimension characteristic analysis on multiple systems to construct a high-dimensional feature matrix for the model input. The correlation between the major features ranges from [0,1]. When the value is less than 0.3, it indicates a weak correlation between the two; when the value is greater than 0.7, it indicates a strong correlation between the two. (4) Construction of input feature set Based on MIC (Multidimensional Correlation Analysis), multidimensional correlation analysis is performed on high-dimensional data. The following data features are used as input data, with a granularity of 15 min: Input1: Predicted cooling load one year prior to the base date; Input2: Predicted heat load one year before the base date; Input3: Predicted electrical load one year prior to the base date; Input4: Predict temperature data for one year prior to the baseline date; Input5: Dew point data predicted one year prior to the base date; Input6: Predicted building data for cooling one year prior to the base date; Input7: Predicted building heat consumption data one year prior to the base date; Input8: Predict building electricity consumption data one year prior to the base date.

6. A multi-element load forecasting device for an integrated energy system, wherein the device employs the steps of the method described in any one of claims 1-5, characterized in that, The device includes: a high-dimensional data receiving module, a high-dimensional data preprocessing module, a multi-dimensional feature analysis module, a data feature mining module, a multi-dimensional feature refinement and extraction module, and a data information prediction and output module. The high-dimensional data receiving module collects data sets from the measurement device and builds a high-dimensional data matrix based on the split window technology, with the split window sliding at intervals. The high-dimensional data preprocessing module uses the Lonely Forest algorithm and the Dynamic Track method to remove outliers and abnormal values ​​in the time series. The multi-dimensional feature analysis module uses the MIC analysis method to perform correlation analysis on the cooling load, heating load, electrical load and meteorological data of the dataset, and select input features; The data feature mining module uses the temporal convolutional unit in the TCN network layer to mine and extract temporal and spatial features from the time series data. The multi-dimensional feature refinement extraction module uses the memory units in the LSTM network layer to refine the input data of the previous layer and mine the long short-term memory relationship. The data information prediction output module utilizes The model's expert network layer and unique gating mechanism predict and output the power values ​​of cooling load, heating load, and electrical load.

7. A computer device for a multi-load forecasting apparatus for an integrated energy system, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images