An energy system load forecasting method based on dynamic hypergraph and time sequence dependence

Abstract

The application provides an energy system load prediction method based on dynamic hypergraph and time sequence dependence, comprising: preprocessing obtained comprehensive energy system historical data to obtain standard input features; processing the standard input features by using a dynamic hypergraph structure learning module to construct a hypergraph structure; extracting high-order correlation features between multivariate load data and meteorological factor data through the hypergraph structure; processing the standard input features by using a Transformer network to extract long time sequence dependence features of the multivariate load data; and adaptively fusing the high-order correlation features and the long time sequence dependence features to obtain multivariate load prediction results. The application suppresses abnormal disturbances through thermal load residual shrinkage, models high-order correlation relationships of multivariate heterogeneous features by using a dynamic hypergraph, and realizes complementation of time sequence dependence and high-order correlation features by using a parallel fusion framework, thereby effectively improving the multivariate load prediction accuracy of the comprehensive energy system.

Description

Technical Field

[0001] This invention relates to the field of integrated energy system load forecasting technology, specifically to an energy system load forecasting method based on dynamic hypergraph and time-series dependence, which is particularly suitable for the joint forecasting of electrical load, cooling load and heating load in integrated energy systems. Background Technology

[0002] Against the backdrop of "dual carbon" goals and energy structure transformation, improving the synergistic utilization efficiency of multi-energy systems has become an important research direction in the energy field. Integrated energy systems (IES), by coupling energy networks such as electricity, heat, and natural gas, can achieve complementary conversion and coordinated supply of multiple energy forms such as electricity, heat, cooling, and gas. Compared with traditional single-energy systems, integrated energy systems have significant advantages in operation regulation, energy cascade utilization, and renewable energy consumption, and therefore have attracted widespread attention from academia and engineering fields.

[0003] Load forecasting is a fundamental aspect of integrated energy system operation optimization and energy management. Accurately predicting the changing trends of electricity, cooling, and heating loads helps the system to formulate scheduling plans in advance, coordinate the operation of energy conversion equipment and energy storage units, and reduce operating costs while meeting user needs. However, the diverse loads of integrated energy systems are affected by a variety of factors, including meteorological conditions, user behavior, equipment operating status, and multi-energy conversion processes, exhibiting obvious nonlinear, non-stationary, and coupled correlation characteristics. In particular, there are differences in energy demand among electricity, cooling, and heating loads, and they are also driven by external environmental factors. Therefore, multi-load forecasting not only needs to characterize the time series variation patterns, but also needs to fully explore the complex coupling relationships between different loads and influencing factors.

[0004] Early research on multi-variable load forecasting mainly followed the modeling approach of single-electricity load forecasting and extended it to integrated energy system scenarios. Representative methods included Kalman filtering, time series analysis, support vector machines, random forests, and gradient boosting trees. These methods are characterized by relatively simple structures and strong interpretability, and can characterize load change trends to a certain extent. However, in integrated energy systems, electrical, cooling, and heating loads are not independent of each other. Their changes are influenced by multi-energy conversion, user behavior, and external environmental factors, exhibiting significant coupling relationships and dynamic dependencies. Traditional methods usually focus on univariate or shallow feature fitting, making it difficult to fully describe the coordinated change patterns among multiple loads. Therefore, the prediction accuracy remains limited in complex operating scenarios.

[0005] In addition, with the continuous improvement of data-driven modeling capabilities, deep learning has gradually become an important technical route for load forecasting of integrated energy systems. Existing research has carried out a lot of exploration around models such as Transformer networks, graph convolutional networks (GCN), temporal convolutional networks (TCN), long short-term memory neural networks (LSTM), and gated recurrent units (GRU). The above-mentioned research has effectively improved load forecasting performance by enhancing the ability of temporal dependency modeling, local feature extraction, and multivariate association expression.

[0006] In terms of correlation modeling, existing research employs methods such as graph convolutional networks combined with K-means clustering, and multi-scale spatiotemporal graph convolutional networks combined with Transformers. By constructing graph structures and utilizing adjacency matrices for information propagation, these methods characterize the spatiotemporal dependencies between multi-node loads, thereby improving load forecasting accuracy. However, these methods are all based on ordinary graph structures for modeling. Since ordinary graph structures mainly rely on edges connecting pairs of nodes to represent the correlation between features, they are difficult to fully explore the heterogeneous multivariate high-order coupling relationships between electricity, cooling, and heating loads and meteorological factors such as temperature and dew point. At the same time, the fixed graph topology is not adaptable enough to the dynamic changes in the coupling relationship between loads and features in integrated energy systems. Therefore, in complex scenarios, there is still a problem of insufficient correlation modeling.

[0007] Furthermore, existing combined prediction models typically integrate decomposition methods, convolutional networks, recurrent networks, or attention mechanisms in a serial cascade manner to enhance the feature representation capabilities of load sequences. For example, existing research has used a TCN–GCN–LSTM serial framework to sequentially complete non-stationary decomposition, temporal feature extraction, coupled feature mining, and long-term dependency modeling, thereby improving the accuracy of multivariate load prediction. Although this serial modeling approach can leverage the advantages of different network modules, its feature transfer mainly relies on layer-by-layer cascaded paths. When multi-scale temporal information and coupled features are modeled together, the serial structure is easily limited by factors such as long information transfer links, insufficient cross-module interaction, and inadequate feature fusion.

[0008] In summary, existing multi-element load forecasting methods for integrated energy systems still have the following technical problems: First, the suppression of random disturbances in heat load is insufficient, meaning that extreme fluctuations and high noise exist in the heat load data, interfering with the stability of the forecasting model; Second, the high-order coupling relationships of multi-element heterogeneous features are not fully characterized, and existing ordinary graph structures cannot fully explore the unpaired high-order correlations between multi-element features; Third, the feature fusion efficiency of serial structures is limited, and the serial cascade method suffers from long information transmission links and insufficient cross-module interaction.

[0009] Therefore, this invention proposes an energy system load forecasting method based on dynamic hypergraph and time-series dependence to solve the above-mentioned technical problems. Summary of the Invention

[0010] The present invention aims to solve the technical problems existing in the prior art and provide an energy system load forecasting method based on dynamic hypergraph and time-series dependence, so as to fully capture the characteristics and changing patterns of IES multivariate load data and improve the technical effect of model prediction.

[0011] To achieve the above objectives, this invention provides an energy system load forecasting method based on dynamic hypergraphs and time-series dependencies, comprising the following steps:

[0012] Step 1, Data Preprocessing: Preprocess the acquired historical data of the integrated energy system to obtain standard input features; specifically:

[0013] Historical data for integrated energy systems includes multi-variable load data and meteorological factor data;

[0014] Multi-variable load data includes electrical load data, cooling load data, and heating load data.

[0015] The preprocessing specifically includes the following steps:

[0016] Step 1a: Decompose the heat load data in the historical data to obtain the heat load residual components; specifically, the decomposition process adopts the STL algorithm, which can decompose the heat load sequence into trend components, periodic components and residual components.

[0017] Step 1b: Apply shrinkage treatment to the residual component of heat load to suppress extreme fluctuations in its residual component.

[0018] Because heat load is affected by factors such as the time lag in heat pipe network transmission, the randomness of user heating behavior, and fluctuations in heat source supply, the fluctuation amplitude in its residual components is significantly higher than that of electrical load and cooling load. Therefore, it is necessary to perform shrinkage processing on the heat load residuals, which can effectively reduce the impact of abnormal disturbances on subsequent feature extraction and improve the quality of the input sequence; specifically:

[0019] The shrinkage process employs a threshold truncation method based on the absolute deviation of the median, including:

[0020] Calculate the median of the absolute values ​​of the residuals of the heat load residual components, and use it as the first statistic.

[0021] Calculate the absolute deviation of the absolute value of each heat load residual relative to the median of the first statistic, and use it as the second statistic.

[0022] The shrinkage threshold is set as a preset multiple of the second statistic.

[0023] The portion of the heat load residual component that is greater than the shrinkage threshold is truncated to the shrinkage threshold, while the portion of the heat load residual component that is less than the shrinkage threshold remains unchanged.

[0024] Step 1c, Feature Selection Processing: The purpose of this step is that the selection of input features is crucial for multivariate load forecasting. Too few input variables will reduce the reliability of the forecast results, while too many variables will increase the computational burden and model uncertainty, thereby affecting the forecast accuracy. Therefore, effectively identifying key influencing factors is a crucial step to ensure the accuracy and reliability of the forecast model.

[0025] Step 1c uses the Pearson correlation coefficient to analyze the correlation, including the following steps:

[0026] First, calculate the Pearson correlation coefficient between meteorological factor data and multivariate load data in historical data; since the Pearson correlation coefficient ranges from -1 to 1, the closer the absolute value is to 1, the stronger the linear correlation between the two variables.

[0027] Then, based on the magnitude of the Pearson correlation coefficient, meteorological factors that are strongly correlated with multivariate loads are selected as input features for the model.

[0028] Specifically, this invention selects temperature, surface albedo, precipitable water, and dew point as strongly correlated factors.

[0029] Step 2, Higher-order correlation feature extraction: The standard input features are processed using the dynamic hypergraph structure learning module to construct a hypergraph structure. The higher-order correlation features between the multivariate load data and meteorological factor data are then extracted using this hypergraph structure. Specifically, this includes the following steps:

[0030] Step 2a: Map the various features in the standard input features to hypergraph nodes respectively. This invention sets up a total of 7 nodes, which correspond to electrical load, cooling load, heating load, temperature, dew point, surface albedo and precipitation.

[0031] Step 2b: Construct and initialize the hypergraph association matrix using the learnable hyperedge weight matrix to build the initial hypergraph structure, which represents the initial association strength between hypergraph nodes and hyperedges.

[0032] This invention does not employ a manually constructed fixed hypergraph structure, but rather adaptively constructs and initializes the hypergraph association matrix using learnable parameters; specifically, it introduces a learnable hyperedge weight matrix. ,in, It is a real matrix. The number of superedges. The number of nodes is given; and the weights are mapped to the 0~1 interval using the Sigmoid activation function to obtain the initial hypergraph structure.

[0033] The initialization of the hypergraph incidence matrix is ​​achieved through the following formula:

[0034]

[0035] In the formula, For the initialized hypergraph incidence matrix, The larger the value of the element, the stronger the connection between the corresponding node and the hyperedge; Use the Sigmoid activation function; is the transpose of the learnable hyperedge weight matrix.

[0036] This invention sets up three potential hyperedges to carry different types of high-order coupling modes between multivariate loads and meteorological factors. The node affiliation of each hyperedge is not pre-set by humans, but is adaptively learned by the model through the learnable hypergraph correlation matrix. Therefore, different hyperedges can be regarded as implicit expressions of potential high-order coupling relationships of multivariate heterogeneous features.

[0037] The construction of the hypergraph structure also includes hyperedge denoising. Considering that there may be weakly associated hyperedges and redundant connections in the initial hypergraph, this invention adopts a sampling mechanism based on Gumbel-Softmax distribution to adaptively retain key hyperedges.

[0038] The super-edge denoising process includes the following steps:

[0039] First, calculate the importance parameter for each hyperedge. , which is a learnable parameter used to represent the importance of the corresponding hyperedge;

[0040] Then, importance parameters are sampled based on the Gumbel-Softmax distribution to generate a hyperedge-preserving mask. ;

[0041] Next, the hyperedges in the initial hypergraph structure are filtered using a hyperedge preservation mask to obtain the denoised hypergraph structure.

[0042] The superedge-preserving mask is generated using the following formula:

[0043]

[0044] In the formula, Preserve the mask for the super-edge; This is a Softmax sampling mechanism based on the Gumbel distribution; Preserve the probability for the superedge; Temperature coefficient; The hard sampling strategy discretizes the output into 0 / 1 form, while using a pass-through estimator (STE) to achieve gradient propagation.

[0045] The Gumbel-Softmax sampling mechanism enables differentiable sampling from discrete distributions and allows for discrete hyperedge selection during training. Its sampling process also incorporates a temperature coefficient... This is used to control the smoothness of the sampling results. The lower the temperature coefficient, the closer the sampling results are to discrete single heat vectors.

[0046] The denoised hypergraph structure is generated using the following formula:

[0047]

[0048] In the formula, The denoised hypergraph structure; This is the initial hypergraph incidence matrix, used to represent the intensity of the association between nodes and hyperedges; This is an element-wise multiplication operation; It is a column vector of length N consisting of all 1s, used to expand the hyperedge mask to the same dimension as the hypergraph incidence matrix.

[0049] By using mask expansion and element-wise filtering operations, invalid or weakly correlated superedges can be effectively removed, while key high-order coupling relationships are preserved.

[0050] The denoised hypergraph structure is then subjected to time-series dynamic update processing and smooth fusion to further express the time-varying characteristics of the coupling relationship between load and meteorological factors.

[0051] Specifically, the time-series dynamic update process includes the following steps:

[0052] First, obtain the historical hypergraph structure from the previous time step. ;

[0053] Then, the temporal fusion weights are calculated. It is used to control the fusion ratio of historical information and current information;

[0054] Finally, based on the temporal fusion weights , historical hypergraph structure The dynamic hypergraph structure is obtained by weighted fusion with the denoised hypergraph structure at the current time step.

[0055] The dynamic hypergraph structure is generated using the following formula:

[0056]

[0057] In the formula, For the updated dynamic hypergraph structure; For time-series fusion weights; Cache the hypergraph from the previous time step; This is the denoised hypergraph structure.

[0058] The time-series dynamic update processing mechanism can improve the adaptability of the hypergraph structure to time-varying high-order coupling relationships while preserving historical association information.

[0059] After obtaining the dynamic hypergraph structure, a two-level aggregation method is used to extract high-order correlation information between multivariate heterogeneous features, that is, to extract high-order correlation features between multivariate load data and meteorological factor data; specifically, the following steps are included:

[0060] First, the input sequence is subjected to temporal global average pooling to obtain global static features. Temporal global average pooling can compress the information of the entire time series into a global feature vector, preserving the overall features of the sequence.

[0061] The global static features are generated using the following formula:

[0062]

[0063] In the formula, It is a global static feature; The length of the input sequence; Let represent the node feature matrix at time step t, where It is a real matrix. For training batch size, This represents the number of nodes.

[0064] Then, based on the dynamic hypergraph structure, global static features are aggregated from hypergraph nodes to hyperedges to obtain hyperedge-level aggregated features, so as to realize the information transfer from nodes to hyperedges, enabling each hyperedge to aggregate the information of all nodes it connects to.

[0065] The generation of the hyperedge-level aggregation feature is achieved through the following formula:

[0066]

[0067] In the formula, It is a hyperedge-level aggregation feature used to encode high-order coupling information between multiple nodes.

[0068] Next, the hyperedge-level aggregated features are back-aggregated from the hyperedge to the hypergraph nodes to obtain global high-order features. This process enables information transfer from the hyperedge to the node, allowing each node to aggregate information from all its associated hyperedges, thereby integrating high-order coupling relationships.

[0069] The generation of the global high-order features is achieved through the following formula:

[0070]

[0071] In the formula, It is a global high-order feature.

[0072] By using the two-level aggregation method described above, from node to hyperedge and then back to node, it is possible to fully extract high-order correlation features between diverse heterogeneous features.

[0073] Finally, the global high-order features are input into a multilayer perceptron network, and nonlinear mapping and dimensionality transformation are performed to obtain the high-order correlation features output by the dynamic hypergraph.

[0074] Step 3: Long-term time-series dependency feature extraction: The Transformer network is used to process the standard input features to extract long-term time-series dependency features from the multivariate workload data; this includes the following steps:

[0075] Step 3a: Map the standard input features to the model dimension of the Transformer network through linear projection. Linear projection is a fully connected layer used to transform the input features from the original dimension to the hidden dimension that the Transformer can handle.

[0076] Step 3b: Inject positional encoding information into the mapped data. Since the Transformer network itself does not have the ability to process sequence order, it is necessary to inject additional positional encoding to preserve temporal information. Positional encoding can be generated using sine and cosine functions, or it can be a learnable positional embedding vector.

[0077] Step 3c: Process the data encoded at the injected positions using stacked Transformer encoders; specifically, the Transformer encoder includes a multi-head attention mechanism and a feedforward network; wherein:

[0078] Multi-head attention is used to adaptively compute attention weights for features at different time steps to extract long-term time-dependent features. Multi-head attention allows the model to focus on information from different locations in different representation subspaces.

[0079] Feedforward networks are used to perform nonlinear transformations on the output of the attention mechanism, thereby enhancing the expressive power of the model.

[0080] After processing by a multi-layer Transformer encoder, long and short-term temporal dependencies in the sequence can be captured. Finally, the aggregated high-order temporal features are decoded through a fully connected network to obtain long-term temporal dependency features.

[0081] Step 4, Fusion Prediction: Adaptively fuse high-order correlation features with long-term temporal dependency features to obtain multivariate load prediction results; where:

[0082] To simultaneously consider the long-term temporal dependencies and high-order coupling relationships between diverse heterogeneous features of load data, this invention constructs a parallel fusion framework of Transformer and dynamic hypergraph. Specifically, the Transformer branch is responsible for modeling the temporal dependencies of historical load sequences, while the dynamic hypergraph branch is used to extract high-order correlation features between load and meteorological factors. In the output stage, learnable fusion weights are introduced to adaptively weight the results of the two branches. The specific process includes the following:

[0083] First, set learnable fusion weight parameters;

[0084] Then, the fusion weight parameters are mapped to the 0~1 range through the Sigmoid activation function to obtain the hypergraph branch weight coefficients, which are used to dynamically adjust the contribution ratio of the hypergraph prediction results to the final prediction results.

[0085] Next, a weighted summation strategy is adopted to fuse higher-order correlation features and long-term time-series dependency features based on the hypergraph branch weight coefficients, so as to obtain the final multivariate load prediction results.

[0086] The fusion is achieved through the following formula:

[0087]

[0088] In the formula, This represents the model's final prediction result; It is a long-term time-dependent feature; For learnable fusion weight parameters; These are the hypergraph branch weight coefficients; This is a high-order correlation feature.

[0089] The above fusion method can dynamically adjust the contribution of the two types of features to the prediction results according to the training process, and achieve complementary fusion of long-term dependent features and high-order correlation features.

[0090] The fusion prediction step also includes inverse normalization. Since the input data is normalized during the data preprocessing stage, compressing the data range to between 0 and 1, the prediction result output by the model is also within the range of 0 to 1. Inverse normalization restores the normalized prediction result to its original dimensions, resulting in a load prediction value with actual physical meaning.

[0091] The beneficial effects of this invention are as follows:

[0092] 1. This invention effectively suppresses extreme fluctuations and high noise disturbances in the residual components by performing seasonal and trend decomposition and residual shrinkage processing on heat load based on local weighted regression, thereby providing a more stable data foundation for subsequent feature extraction and model prediction, and thus improving the prediction stability of the model.

[0093] 2. This invention constructs a dynamic hypergraph structure learning module, which maps multiple loads and meteorological factors into hypergraph nodes in a unified manner. It uses hyperedges to describe the high-order coupling relationship between multiple variables, which can effectively characterize the time-varying high-order correlation between electricity, cooling and heating loads and meteorological factors, making up for the shortcomings of ordinary graph structures that can only model binary relationships.

[0094] 3. By introducing a super-edge denoising mechanism based on the sampling mechanism of Gumbel-Softmax distribution and a time-series dynamic update mechanism, this invention can adaptively filter out effective coupling relationships and capture the dynamic evolution characteristics of coupling relationships, thereby enhancing the model's ability to represent time-varying high-order correlation features.

[0095] 4. This invention designs a parallel fusion framework of Transformer and dynamic hypergraph. The Transformer branch extracts long-term time-dependent features, and the dynamic hypergraph branch extracts high-order correlation features. The two types of features are dynamically complementary through adaptive weight fusion, which overcomes the problems of limited feature fusion and long information transmission links in serial structure, thereby improving the accuracy of multivariate load prediction. Attached Figure Description

[0096] Figure 1 This is a schematic diagram of the process of the present invention;

[0097] Figure 2 This is a schematic diagram comparing the electrical load prediction of this invention with other prediction models.

[0098] Figure 3 This is a schematic diagram comparing the cooling load prediction of this invention with other prediction models.

[0099] Figure 4 This is a schematic diagram comparing the heat load prediction of an embodiment of the present invention with other prediction models. Detailed Implementation

[0100] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0101] Example

[0102] This embodiment provides an energy system load forecasting method based on dynamic hypergraph and time-series dependency, which is used to integrate the energy system and simultaneously forecast electrical load, cooling load and heating load.

[0103] The data used in this embodiment comes from detailed load data provided by a university, which covers electricity, cooling, and heating load data for one and a half years from 0:00 on January 1, 2021 to 23:00 on December 31, 2023, with a time resolution of 1 hour. The meteorological factors include 9 categories: temperature, surface albedo, wind direction, wind speed, pressure, relative humidity, precipitable water, dew point, and cloud type. At the same time, time, day, and month are also considered as input features.

[0104] Please refer to Table 1. This embodiment is based on a parallel prediction framework. The training hyperparameters are set as follows: 30 training iterations, 32 training batches, 24 time window length, L1Loss loss function, Adam optimizer, 0.0005 initial learning rate, 1e-6 weight decay coefficient, and 5 patience value for the learning rate scheduler.

[0105] Table 1 Model Parameter Settings

[0106]

[0107] To train and validate the performance of the model in this embodiment, the training set, validation set, and test set are divided in a ratio of 8:1:1; the prediction step is set to 1 hour, and the data is used in a rolling prediction manner.

[0108] This embodiment achieves prediction through the following steps:

[0109] First, perform data preprocessing steps; acquire historical data of the integrated energy system, including multi-variable load data and meteorological factor data; specifically:

[0110] Multi-variable load data includes electrical load data, cooling load data, and heating load data.

[0111] Meteorological data include temperature, surface albedo, wind direction, wind speed, pressure, relative humidity, precipitable water, dew point, and cloud type.

[0112] Preprocessing refers to outlier detection and processing of the raw data; specifically, the 3σ criterion is used for outlier detection. The 3σ criterion assumes that the data follows a normal distribution and considers data points that exceed the mean plus or minus three standard deviations as outliers.

[0113] For detected outliers, a linear interpolation method between two adjacent valid sampling points is used to fill and replace them.

[0114] Calculate the Pearson correlation coefficient between meteorological factor data and multivariate load data in historical data.

[0115] Based on the calculated Pearson correlation coefficient, meteorological factors with strong correlation to multivariate loads were selected as model input features. Table 2 shows the Pearson correlation coefficients between multivariate loads and various weather factors.

[0116] Table 2. Pearson correlation coefficients between multiple loads and weather factors.

[0117]

[0118] As shown in Table 2, the degree of influence of meteorological factors on different loads varies. For example, the correlation coefficients between relative humidity and electricity, cooling, and heating loads are -0.074, -0.075, and 0.059, respectively. At the same time, there are strong correlations between multiple loads and between loads and some meteorological factors. For example, the correlation coefficient between electricity load and cooling load is 0.754, and the correlation coefficient between temperature and cooling load is 0.715. It can be seen that there is a cross-type high-order correlation and synergistic change law between loads and meteorological factors in the integrated energy system.

[0119] Therefore, in this embodiment, four sets of weather data with high correlation coefficients are selected as strong correlation factors for prediction: temperature, surface albedo, precipitable water and dew point. As can be seen from Table 2, the correlation coefficients between the selected data and the electricity, cooling and heating loads are all greater than 0.4. Here, 0.4 refers to the absolute value of the Pearson correlation coefficient.

[0120] Processing heat load data involves the following steps:

[0121] First, the heat load data is decomposed using the STL algorithm. Specifically, the STL algorithm is a time series decomposition method that can decompose a time series into a trend component, a periodic component, and a residual component. The trend component is used to reflect the long-term direction of change in the series, the periodic component is used to reflect the recurring seasonal patterns in the series, and the residual component is used to remove the trend and periodic components, which contain random noise and abnormal fluctuations.

[0122] Extensive experimental data show that there are more significant extreme fluctuations in the residual components of heat load. Therefore, this embodiment further performs shrinkage processing on the residual components of heat load, which is based on the median absolute deviation. As a robust statistical measure, the median absolute deviation is not sensitive to outliers. The shrinkage processing reduces the residuals that exceed a certain multiple of the median absolute deviation, thereby suppressing extreme fluctuations in the residual components and obtaining a new data sequence after processing.

[0123] Finally, to improve the convergence speed and accuracy of the model, Min-Max normalization was performed on all features input to the model and the processed multivariate loading data. Normalization compresses the data range to between 0 and 1, eliminating the influence of different dimensions between features; specifically:

[0124] The normalization is achieved through the following formula:

[0125]

[0126] In the formula, This is the original data; and These represent the minimum and maximum values ​​of the feature in the dataset, respectively. This is the result after normalization.

[0127] After data preprocessing is completed, the high-order correlation feature extraction step and the long temporal dependency feature extraction step are executed in parallel.

[0128] In the higher-order correlation feature extraction step, dynamic hypergraphs are used to process the preprocessed data.

[0129] Dynamic hypergraphs include hypergraph initialization, hyperedge denoising, temporal dynamic updates, and feature aggregation; among which:

[0130] Hypergraph initialization is used to construct the initial hypergraph structure, mapping various features in the preprocessed data to hypergraph nodes. In this embodiment, a total of 7 nodes are set, corresponding to electrical load, cooling load, heating load, temperature, dew point, surface albedo, and precipitation.

[0131] Instead of using a manually constructed fixed hypergraph structure, the hypergraph initialization introduces a learnable hyperedge weight matrix. The weight matrix is ​​then mapped to the 0-1 interval using the Sigmoid activation function to obtain the initial hypergraph association matrix.

[0132] The above process is achieved through the following formula:

[0133]

[0134] This embodiment sets up three potential hyperedges to carry different types of higher-order coupling patterns. The node membership relationships of each hyperedge are not pre-defined, but are adaptively learned by the model through the learnable hypergraph association matrix. Therefore, different hyperedges can be regarded as implicit expressions of potential higher-order coupling relationships of multi-dimensional heterogeneous features.

[0135] Hyperedge denoising refers to filtering hyperedges in the initial hypergraph structure. The initial hypergraph may contain weakly correlated hyperedges and redundant connections. These weakly correlated hyperedges will introduce noise and interfere with the model's prediction performance.

[0136] Hyperedge denoising employs a hyperedge selection method based on a sampling mechanism using the Gumbel-Softmax distribution, adaptively retaining key hyperedges; specifically:

[0137] First, this embodiment sets up a total of 3 hyperedges. For each hyperedge, a learnable importance parameter is defined. The importance parameters of each hyperedge together form a vector α of length 3, where α1, α2, and α3 correspond to the importance of the 1st, 2nd, and 3rd hyperedges, respectively.

[0138] Then, the hyperedge retention probability is obtained through the Sigmoid activation function;

[0139] Next, the above retention probabilities are sampled using the Gumbel-Softmax distribution to generate discrete hyperedge retention masks.

[0140] This process is achieved through the following formula:

[0141]

[0142] The sampling process based on Gumbel-Softmax is as follows:

[0143] Noise is sampled from the Gumbel distribution and normalized exponentially; the temperature coefficient controls the smoothness of the sampling results, and in this embodiment, the temperature coefficient is set to 0.4.

[0144] The generated mask is multiplied element-wise with the initial hypergraph association matrix to obtain the denoised hypergraph structure.

[0145] The above process is achieved through the following formula:

[0146]

[0147]

[0148] The above operations can filter out invalid or weakly correlated superedges, whose corresponding mask value is 0; and retain key high-order coupling relationships, whose corresponding mask value is 1.

[0149] Time-series dynamic updates are used to capture the time-varying characteristics of the coupling relationship between load and meteorological factors, because the coupling relationship in an integrated energy system is not static but evolves dynamically over time; specifically:

[0150] First, obtain the cached historical hypergraph structure from the previous time step and calculate a learnable temporal fusion weight:

[0151]

[0152] In the formula, Learnable weight scalars for temporal fusion.

[0153] In this embodiment, the initial value of the temporal fusion weight is set to 0.5.

[0154] Then, based on the temporal fusion weight, the historical hypergraph structure and the denoised hypergraph structure output by the hyperedge denoising at the current time are weighted and summed to obtain the updated dynamic hypergraph structure, which is achieved by the following formula:

[0155]

[0156] This dynamic hypergraph structure retains historical association information while adapting to current changes in coupling relationships.

[0157] Feature aggregation refers to extracting high-order association features based on the updated dynamic hypergraph structure; this embodiment adopts a two-level aggregation method of nodes, hyperedges, and nodes; specifically:

[0158] First, the input time series is subjected to time-series global average pooling to obtain global static features, which is achieved through the following formula:

[0159]

[0160] In the formula, Let represent the node feature matrix at time step t, where It is a real matrix. For training batch size, In this embodiment, the number of nodes is... The value is 7; T is the time step, which is 24 in this embodiment.

[0161] Then, based on the dynamic hypergraph structure, the global static features are aggregated from the hypergraph nodes to the hyperedges to obtain the hyperedge-level aggregated features, which is achieved through the following formula:

[0162]

[0163] In the formula, This represents a hyperedge-level aggregation feature used to encode higher-order coupling information between multiple nodes, where... It is a real matrix. For training batch size, In this embodiment, the number of superedges is... The value is 3.

[0164] Next, the hyperedge-level aggregated features are back-aggregated from the hyperedges back to the hypergraph nodes to obtain global high-order features, which is achieved through the following formula:

[0165]

[0166] In the formula, These are global high-order features, among which, It is a real matrix. For training batch size, This represents the number of nodes.

[0167] Finally, the aforementioned global high-order features are input into a multilayer perceptron network, and nonlinear mapping and dimensional transformation are performed to obtain the high-order correlation features output by the dynamic hypergraph.

[0168] In this embodiment, the multilayer perceptron network includes two fully connected layers.

[0169] In the long-term time-dependent feature extraction step, the Transformer network is used to process the preprocessed data. The Transformer network is used to model the long-term time-dependent relationships of historical load sequences.

[0170] After processing by the Transformer encoder, the model is able to capture long and short-term temporal dependencies in the sequence.

[0171] The aggregated high-order temporal features are decoded by a fully connected network to obtain the long temporal dependency features output by the Transformer network. The output dimension is 3, which corresponds to the predicted values ​​of electricity, cooling and heating loads, respectively.

[0172] After obtaining the high-order association features from the dynamic hypergraph output and the long-term temporal dependency features from the Transformer network output, a fusion prediction step is performed, which is completed by the adaptive fusion module.

[0173] The adaptive fusion module sets a learnable fusion weight parameter, which is mapped to the 0~1 range through the Sigmoid activation function to obtain the hypergraph branch weight coefficient. This coefficient is used to dynamically adjust the contribution ratio of the hypergraph prediction result.

[0174] The adaptive fusion module employs a weighted summation strategy to fuse the long-term temporal dependency features and higher-order association features output by the Transformer network according to the aforementioned weight coefficients. The fusion formula is as follows:

[0175]

[0176] The above fusion method can dynamically adjust the contribution of the two types of features to the prediction results according to the training process, and achieve complementary fusion of long-term dependent features and high-order correlation features.

[0177] Finally, the adaptive fusion module outputs the fused predictions for electrical load, cooling load, and thermal load.

[0178] Since Min-Max normalization was performed during the data preprocessing stage, the prediction results need to be denormalized here to restore them to their original dimensions.

[0179] To verify the effectiveness of this embodiment, ablation experiments and control experiments were set up; among them, four control experiments were designed in the ablation experiment:

[0180] The first group did not perform residual shrinkage on the heat load, but the dynamic supergraph module and the Transformer module were enabled;

[0181] The second group is the complete model, which simultaneously enables the residual shrinkage, dynamic hypergraph module, and Transformer module;

[0182] The third group enabled residual shrinkage and the Transformer module but disabled the dynamic hypergraph module;

[0183] The fourth group is the baseline model, with only the Transformer module enabled, without residual shrinkage and dynamic hypergraph modules.

[0184] The above model uses a weighted average absolute percentage error (... As evaluation indicators, the weighting coefficients for electricity, cooling, and heating loads are 0.5, 0.3, and 0.2, respectively.

[0185] Table 3 Ablation Experiment Results

[0186]

[0187] The above experimental results show that the model in this embodiment (group 2) performs well. The results were superior to those of other experimental groups, validating the effectiveness of the combined use of the dynamic hypergraph module and the residual shrinkage strategy.

[0188] Furthermore, the dynamic hypergraph-Transformer model of this embodiment is compared with the static hypergraph-Transformer model.

[0189] Table 4 Comparison of computational efficiency of different models

[0190]

[0191] As shown in Table 4, the weighted average absolute percentage error of the dynamic model is 2.120%, and the weighted root mean square error is 281.958, which is better than the static model's 2.154% and 288.019, respectively. The test time only increases by 0.042 seconds, and the single-sample inference time only increases by 0.017 milliseconds, indicating that the dynamic update mechanism can improve prediction accuracy at an acceptable time cost.

[0192] The following comparative experiments compare the model in this embodiment (hereinafter referred to as: Dynamic Hypergraph-Transformer-MTL) with the CNN-BiLSTM multi-task learning model (hereinafter referred to as: CNN-BiLSTM-MTL), the GRU-Transformer multi-task learning model (hereinafter referred to as: GRU-Transformer-MTL), the graph attention network multi-task learning model (hereinafter referred to as: GAT-MTL), and the Dynamic Hypergraph-Transformer single load prediction model (hereinafter referred to as: single load prediction model) to simultaneously predict the electricity, cooling, and heating loads within a certain time period, and then compare and analyze them with the actual values ​​of electricity, cooling, and heating loads.

[0193] Specifically, a certain time period refers to a consecutive week selected from the test set (from 12:00 on September 14, 2023 to 11:00 on September 21, 2023) as the time period to demonstrate the load forecasting effect.

[0194] Table 5. Error Analysis of Different Models

[0195]

[0196] In Table 5, This is the root mean square error; The mean absolute percentage error.

[0197] Figure 2 The unit of electrical load value is kW. Figure 3 The unit for intercooling load is Ton. Figure 4 The unit of medium heat load is MMBtu / h, and 1kW = 0.284Ton = 0.0034MMBtu / h.

[0198] From Table 5 and Figures 2-4 This indicates that the model in this embodiment has an average absolute percentage error in terms of electrical, cooling, and heating loads ( The percentages were 1.99%, 2.45%, and 1.95%, respectively, with a weighted average absolute percentage error (...). The weighted average absolute percentage error was 2.120%, which is better than other comparative models. Compared with the graph attention network multi-task learning model, the weighted average absolute percentage error was reduced by 0.511%. Compared with the GRU-Transformer multi-task learning model, the weighted average absolute percentage error was reduced by 0.610%.

[0199] Therefore, it can be seen that the heat load residual shrinkage strategy, dynamic hypergraph structure learning module, and parallel fusion framework adopted in this embodiment can effectively improve the prediction accuracy, and the modules have good complementarity; specifically:

[0200] Compared with single-load forecasting models, the multi-load joint forecasting model in this embodiment can effectively capture the coupling effect between different loads, reduce information silos, and thus obtain more comprehensive and reliable forecasting results. In addition, by sharing features and information in the same model, data resources can be utilized more efficiently, further improving the accuracy and efficiency of forecasting. Furthermore, through comparative analysis of single-load forecasting and multi-load forecasting results, it is concluded that joint forecasting of electricity, cooling, and heating loads is a more effective and accurate load forecasting method, which can significantly improve forecasting accuracy.

[0201] Compared with GAT-MTL, the model in this embodiment demonstrates superior performance in all three types of load forecasting. The metric decreased by 0.511%. This embodiment combines the advantages of dynamic hypergraph and Transformer to effectively improve prediction accuracy, further verifying the advantages of the model designed in this embodiment in feature capture and learning.

[0202] Compared with traditional time series prediction models such as CNN-BiLSTM-MTL and GRU-Transformer-MTL, the advantage of the model in this embodiment lies in the synergistic effect of the dynamic hypergraph and the Transformer network: Taking the metrics as an example, compared to GRU-Transformer-MTL, this embodiment reduces the accuracy by 0.610%. The main difference between GRU-Transformer and this embodiment lies in GRU and dynamic hypergraph. GRU excels at time series modeling, adaptively capturing short-term dependencies and long-term trends of time series through reset and update gates, effectively solving the gradient vanishing problem of long time series. Hypergraph, on the other hand, better mines higher-order coupling relationships between features through graph structures. Therefore, the model in this embodiment retains the Transformer's ability to model long-term time series dependencies and adds the ability of dynamic hypergraph to mine higher-order dynamic associations between electrical, cooling, and heating loads and environmental variables, thereby significantly improving prediction accuracy.

Claims

1. A method for load forecasting of energy systems based on dynamic hypergraphs and time-series dependencies, characterized in that, Includes the following steps: The acquired historical data of the integrated energy system is preprocessed to obtain standard input features; specifically, the historical data includes multi-variable load data and meteorological factor data, and the multi-variable load data includes electrical load data, cooling load data, and heating load data. The standard input features are processed using a dynamic hypergraph structure learning module to construct a hypergraph structure. High-order correlation features between multivariate load data and meteorological factor data are then extracted through the hypergraph structure. The Transformer network is used to process the standard input features to extract long-term temporal dependency features of multivariate load data; Adaptive fusion of high-order correlation features and long-term time-series dependency features yields multivariate load prediction results.

2. The energy system load forecasting method according to claim 1, characterized in that, The preprocessing of historical data from the integrated energy system includes the following steps: The heat load data in the historical data is decomposed to obtain the heat load residual components; specifically, the decomposition process adopts the STL algorithm. The residual component of the heat load is contracted to suppress abnormal fluctuations in the residual.

3. The energy system load forecasting method according to claim 1, characterized in that, The construction of the hypergraph structure includes the following steps: Map each type of feature in the standard input features to a hypergraph node; By using a learnable hyperedge weight matrix, a hypergraph association matrix is ​​constructed and initialized to build an initial hypergraph structure. The hypergraph association matrix is ​​used to represent the initial association strength between hypergraph nodes and hyperedges. The initialization of the hypergraph incidence matrix is ​​achieved through the following formula: ； In the formula, For the initialized hypergraph incidence matrix, The larger the value of the element, the stronger the connection between the corresponding node and the hyperedge; Use the Sigmoid activation function; is the transpose of the learnable hyperedge weight matrix.

4. The energy system load forecasting method according to claim 3, characterized in that, The construction of the hypergraph structure also includes hyperedge denoising processing, which includes the following steps: Calculate the importance parameter for each hyperedge; The importance parameter is sampled based on the sampling mechanism of Gumbel-Softmax distribution to generate a hyperedge-preserving mask; The superedge-preserving mask is generated using the following formula: ； In the formula, Preserve the mask for the super-edge; This is a Softmax sampling mechanism based on the Gumbel distribution; Preserve the probability for the superedge; Temperature coefficient; The hard sampling strategy discretizes the output into 0 / 1 form, while the pass-through estimator (STE) is used to implement gradient propagation. The hyperedges in the initial hypergraph structure are filtered using a hyperedge-preserving mask to obtain a denoised hypergraph structure. The denoised hypergraph structure is generated using the following formula: ； In the formula, The denoised hypergraph structure; This is an element-wise multiplication operation; It is a column vector of length N consisting of all 1s, used to expand the hyperedge mask to the same dimension as the hypergraph incidence matrix.

5. The energy system load forecasting method according to claim 4, characterized in that, The denoised hypergraph structure is then subjected to time-series dynamic update processing, which includes the following steps: Obtain the historical hypergraph structure of the previous time step; Calculate the temporal fusion weights; Based on the temporal fusion weights, the historical hypergraph structure and the denoised hypergraph structure at the current moment are weighted and fused to obtain the dynamic hypergraph structure. The dynamic hypergraph structure is generated using the following formula: ； In the formula, For the updated dynamic hypergraph structure; For time-series fusion weights; This is the historical hypergraph structure cached from the previous time step.

6. The energy system load forecasting method according to claim 5, characterized in that, The extraction of high-order correlation features between multivariate load data and meteorological factor data includes the following steps: Temporal global average pooling is performed on the input sequence to obtain global static features, which is achieved by the following formula: ； In the formula, It is a global static feature; The length of the input sequence; Let be the node feature matrix at time step t, and ,in, It is a real matrix. For training batch size, The number of nodes; Based on the dynamic hypergraph structure, global static features are aggregated from hypergraph nodes to hyperedges to obtain hyperedge-level aggregated features, which is achieved through the following formula: ； In the formula, This is a hyper-edge-level aggregation feature; For dynamic hypergraph correlation matrix; The hyperedge-level aggregated features are back-aggregated from the hyperedge to the hypergraph nodes to obtain global high-order features, which is achieved by the following formula: ； In the formula, It is a global high-order feature; This is the transpose of the dynamic hypergraph incidence matrix; Finally, the global high-order features are input into a multilayer perceptron network, and nonlinear mapping and dimensionality transformation are performed to obtain the high-order correlation features output by the dynamic hypergraph.

7. The energy system load forecasting method according to claim 1, characterized in that, The extraction of long-term time-series dependency features from multivariate load data includes the following steps: Standard input features are mapped to the model dimensions of the Transformer network via linear projection; Inject position encoding information into the mapped data; The injected position encoded data is processed using stacked Transformer encoders; The Transformer encoder includes a multi-head attention mechanism and a feedforward network to adaptively calculate the attention weights of features at different time steps in order to extract long-term temporal dependent features.

8. The energy system load forecasting method according to claim 6, characterized in that, The adaptive fusion of high-order correlation features and long-term temporal dependency features includes the following steps: Set learnable fusion weight parameters; The fusion weight parameters are mapped to the 0~1 range by using the Sigmoid activation function to obtain the hypergraph branch weight coefficients; A weighted summation strategy is adopted to fuse higher-order correlation features and long-term time-series dependency features based on the hypergraph branch weight coefficients to obtain the final multivariate load prediction results. The fusion is achieved through the following formula: ； In the formula, This is the final result of the multivariate load forecast; It is a long-term time-dependent feature; Use the Sigmoid activation function; These are learnable fusion weight parameters; These are the hypergraph branch weight coefficients; This is a high-order correlation feature.

9. The energy system load forecasting method according to claim 1, characterized in that, The preprocessing of the historical data of the integrated energy system also includes feature filtering, which includes the following steps: Calculate the Pearson correlation coefficient between meteorological factor data and multivariate load data in the historical data, and select meteorological factors with strong correlation with multivariate load as model input features based on the magnitude of the Pearson correlation coefficient.

10. The energy system load forecasting method according to claim 2, characterized in that, The shrinkage process employs a threshold truncation method based on the absolute deviation of the median, including: Calculate the median of the absolute values ​​of the residuals of the heat load residual components, and use it as the first statistic. Calculate the absolute deviation of the absolute value of each heat load residual relative to the median of the first statistic, and use it as the second statistic. The shrinkage threshold is set as a preset multiple of the second statistic. The portion of the heat load residual component that is greater than the shrinkage threshold is truncated to the shrinkage threshold, while the portion of the heat load residual component that is less than the shrinkage threshold remains unchanged.

Images