A district energy system multi-source load prediction method and system

Abstract

The application provides a kind of district energy system multi-source load prediction method and system, related to the technical field of load prediction, the method comprises: by obtaining the historical multi-source load of district energy system, meteorological data and equipment parameters, the load data is decomposed using maximum overlap discrete wavelet transform, and a coupling correlation graph is constructed with electric cold and heat load as node and equipment parameter as weight to extract topological features;Then build a multi-task LSTM model, combine different frequency domain load components, historical meteorological data and coupling feature vectors for multi-task learning training, finally input meteorological forecast and coupling feature vector to predict each frequency domain component and inverse wavelet transform reconstruction, realize the accurate prediction of electric, cold and heat load.The application can improve the accuracy of multi-source load prediction results.

Description

Technical Field

[0001] This application relates to the technical field of load forecasting, and in particular to a method and system for multi-source load forecasting of regional energy systems. Background Technology

[0002] As a key vehicle for realizing energy cascade utilization and low-carbon transformation, the efficient operation of regional energy systems relies on accurate short-term forecasts of multi-source loads, including electricity, cooling, and heating. Accurate load forecasting not only guides the optimal scheduling of energy equipment and reduces operating costs but also effectively mitigates system risks arising from renewable energy fluctuations. However, the multi-source loads in a regional energy system are not independent but are physically closely linked through coupled devices such as heat pumps, chiller units, and waste heat recovery units, and are heavily influenced by meteorological conditions, user behavior, and special events. This strong coupling, nonlinearity, and multi-timescale characteristic makes it difficult for traditional single-load forecasting methods or simple statistical regression models to capture the complex dynamics within the system.

[0003] Chinese invention patent application publication number CN120511656A, with an application publication date of August 19, 2025, discloses a method for short-term prediction of multiple loads in an integrated energy system. This method utilizes maximum overlap discrete wavelet transform to decompose load sequences and constructs a hybrid model combining a temporal convolutional network and a Transformer. It shares underlying features through a multi-task learning strategy to uncover the implicit correlations between electricity, cooling, and heating loads, while employing an improved energy valley optimization algorithm to optimize the model's hyperparameters.

[0004] While this scheme utilizes the ability of deep learning to extract time-series features to some extent and achieves joint prediction of multiple loads through multi-task learning, its core logic still focuses on statistically mining the correlation between variables from historical data, treating the coupling relationship between multiple source loads as a black box mapping, which makes the prediction model unable to perceive the actual connection mode and energy conversion path of equipment in the regional energy system.

[0005] This means that when the system's operating conditions change drastically (such as significant fluctuations in equipment partial load rate, equipment performance drift, or encountering extreme coupling scenarios not seen in historical data), the model, lacking constraints and guidance from physical mechanisms, is prone to producing prediction results that violate the laws of thermodynamics, leading to a decrease in the accuracy of multi-source load prediction results. Summary of the Invention

[0006] To improve the accuracy of multi-source load forecasting results, this application provides a method and system for multi-source load forecasting in regional energy systems.

[0007] Firstly, this application provides a method for predicting multi-source loads in a regional energy system, employing the following technical solution: A method for predicting multi-source loads in a regional energy system includes the following steps: Historical multi-source load data, historical meteorological data and equipment rated thermodynamic conversion parameters of the regional energy system are obtained. The historical multi-source load data are decomposed by maximum overlap discrete wavelet transform to obtain load components in different frequency domains. Using electrical, cooling, and heating loads as nodes and the rated thermodynamic conversion parameters of the equipment as edge weights, a multi-source load coupling correlation graph is constructed. Topological features are extracted from the multi-source load coupling correlation graph, and coupling feature vectors are constructed based on the topological features. A multi-task LSTM model is constructed and trained using load components, historical meteorological data, and coupled feature vectors in different frequency domains. The load components, meteorological forecast data, and coupled feature vectors corresponding to historical multi-source load data of a preset duration between the current time are input into the trained multi-task LSTM model to obtain the predicted values ​​of each frequency domain component. The predicted values ​​of each frequency domain component are reconstructed by maximum overlap discrete wavelet inverse transform to obtain the predicted values ​​of electricity, cooling, and heating loads.

[0008] This application constructs a multi-source load coupling correlation graph using electrical, cooling, and heating loads as nodes and the rated thermodynamic conversion parameters of equipment as edge weights. Topological features are extracted from this graph, and coupling feature vectors are constructed based on these features. This fully utilizes the coupling relationships between loads to improve prediction accuracy. This application also constructs a multi-task LSTM model. By sharing some model parameters, the multi-task LSTM model can utilize the correlations between different types of loads, improving the model's generalization ability and prediction accuracy.

[0009] This application uses load components in different frequency domains, historical meteorological data, and coupled feature vectors to train a multi-task LSTM model. Load components in different frequency domains provide the characteristics of load at different time scales; historical meteorological data takes into account the impact of external environmental factors on load; and coupled feature vectors reflect the interrelationships between different types of loads. By integrating these features into the model, the multi-task LSTM model can learn the load change patterns from multiple perspectives, thereby improving the accuracy of prediction.

[0010] This application reconstructs the predicted values ​​of electricity, cooling, and heating loads by performing maximum overlap discrete wavelet inverse transform on the predicted values ​​of each frequency domain component. This method can better handle the complex features in the load data, reduce prediction errors, and improve prediction accuracy.

[0011] Optionally, the method further includes: Identify the coupled devices within the regional energy system, including chiller units, heat pump units, and waste heat recovery devices. Construct characteristic curves of the energy efficiency ratio of each coupled device as a function of partial load rate. Establish a mapping relationship between electrical power input and cooling / heating power output based on the characteristic curves. Calculate the marginal contribution rate of a unit change in electrical load to the change in cooling and heating load based on the mapping relationship. Update the weights of corresponding edges in the multi-source load coupling correlation graph using the marginal contribution rate.

[0012] This application first identifies coupled equipment such as refrigeration units, heat pump units, and waste heat recovery devices within the regional energy system, and constructs characteristic curves of the energy efficiency ratio of each coupled equipment as a function of the partial load rate. Then, based on the characteristic curves, it establishes a mapping relationship between electrical power input and cooling / heating power output, further clarifying the quantitative connection between the equipment's input energy and output energy.

[0013] Subsequently, this application calculates the marginal contribution rate of a unit change in electrical load to changes in cooling and heating loads based on the mapping relationship. The marginal contribution rate accurately reflects the impact of small changes in electrical load on cooling and heating loads. Finally, the weights of corresponding edges in the multi-source load coupling correlation graph are updated using the marginal contribution rate. This allows the multi-source load coupling correlation graph to more accurately reflect the actual correlation between different loads in the energy system. The updated multi-source load coupling correlation graph can more realistically simulate the operation of the energy system. Based on the updated correlation graph, the mutual influence between different loads can be considered more accurately, further improving the accuracy of the model's prediction results.

[0014] Optionally, the method further includes: calculating in real time the deviation between the actual energy efficiency ratio of the coupling device and the theoretical energy efficiency ratio obtained based on the characteristic curve; when the deviation continues to exceed the allowable error range within a preset time window, resetting the weight of the corresponding edge of the coupling device in the multi-source load coupling correlation diagram to zero, until the deviation does not exceed the allowable error range.

[0015] This application enables timely detection of changes in equipment operating status by calculating the deviation between the actual energy efficiency ratio (EER) of coupled equipment (refrigeration units, heat pump units, and waste heat recovery devices, etc.) and the theoretical EER obtained based on characteristic curves in real time. The theoretical EER is determined based on the characteristic curves of the equipment during normal operation and represents the equipment's performance under ideal or expected operating conditions. The actual EER, on the other hand, reflects the equipment's true performance under current operating conditions. By comparing the deviation between the two in real time, this application can quickly identify whether the equipment is experiencing abnormal operating conditions.

[0016] In actual operation, the energy efficiency ratio of equipment may fluctuate temporarily due to factors such as instantaneous changes in ambient temperature or brief fluctuations in grid voltage. This application filters out these accidental factors by triggering subsequent operations only when the equipment exceeds the allowable error range within a preset time window, making the diagnosis of equipment faults more accurate and reliable.

[0017] This application ensures that the weights of edges in the multi-source load coupling correlation graph remain consistent with the actual operation of the energy system by adjusting the weights of edges in real time based on equipment status. If the correlation graph fails to reflect the impact of equipment failures on load coupling relationships in a timely manner, subsequent analyses based on the graph will be inaccurate and may lead to erroneous decisions. This application promptly resets the weights of edges corresponding to faulty equipment to zero, effectively preventing the continued consideration of the equipment's energy conversion capacity during energy dispatching and thus preventing incorrect dispatching decisions.

[0018] Optionally, the method further includes: Establish an event library containing event tags for statutory holidays, exhibitions, and sporting events; and construct a corresponding load offset correction coefficient matrix for each event tag. If the prediction period matches any event label, then the Hadamard product operation is performed between the correction coefficient matrix corresponding to the prediction period and the topological features to generate a coupled feature vector containing the event perturbation factor.

[0019] This application establishes an event database containing event tags such as statutory holidays, exhibitions, and sporting events, enabling it to cover various special events that may significantly impact energy load. Special events often disrupt regular energy usage patterns, leading to load variations that differ from normal days. Furthermore, this application constructs a corresponding load offset correction coefficient matrix for each event tag, quantifying the specific impact of different events on energy load. Different types of events affect load in different ways and with varying magnitudes; by constructing a separate correction coefficient matrix for each event tag, this application can more accurately describe these differences, thereby improving the accuracy of load forecasting.

[0020] When the prediction period matches any event label, the correction coefficient matrix corresponding to the prediction period is subjected to Hadamard product operation with the topological features to generate a coupled feature vector containing event disturbance factors. Through Hadamard product operation, the two are organically combined so that the coupled feature vector contains both the inherent topological information of the energy system and the disturbance factors brought about by special events, thereby further improving the prediction accuracy and reducing prediction error.

[0021] Optionally, the method further includes: Obtain the actual measured values ​​of multi-source loads, calculate the residual sequence between the predicted value and the actual measured value at the corresponding time, smooth the residual sequence using the moving average algorithm, predict the residual value at the next time, add the predicted value of the electricity, cooling and heating load at the next time to the residual value at the next time, and obtain the final predicted value of the electricity, cooling and heating load at the next time.

[0022] This application obtains the actual measured values ​​of multi-source loads (electricity, cooling, and heating loads) and calculates the residual sequence between the predicted and actual measured values ​​at corresponding times, enabling precise location of the differences between the predicted and actual values. By calculating the residual sequence, the magnitude and distribution of these deviations can be clearly identified. Subsequently, a moving average algorithm is used to smooth the residual sequence, effectively reducing random fluctuations and noise interference in the residual sequence.

[0023] This application improves prediction accuracy by predicting the residual value at the next time step and adding the predicted values ​​of electricity, cooling, and heating loads at the next time step to the residual value at the next time step. By predicting the residual sequence, this application can estimate the possible deviations in the predicted value at the next time step in advance and correct them into the initial predicted value, thereby obtaining a final prediction result that is closer to the actual value.

[0024] Optionally, the method further includes: The residual standard deviation of each type of load is calculated based on the residual sequence. The confidence interval of each type of load is constructed based on the final predicted value of each type of load and the residual standard deviation of that type of load. The width of the confidence interval is then calculated. The system acquires the minimum stable operating load threshold and the preset uncertainty tolerance threshold for energy equipment within the region. When the interval width exceeds the uncertainty tolerance threshold, or the lower limit of the confidence interval is lower than the minimum stable operating load threshold, an early warning signal is issued.

[0025] This application calculates the residual standard deviation for each type of load (electricity, cooling, heating, etc.) based on the residual sequence. The residual standard deviation directly reflects the error fluctuation of load forecasting. The larger the residual standard deviation, the more drastic the fluctuation between the predicted and actual values, and the higher the forecast uncertainty; conversely, the smaller the residual standard deviation, the stronger the forecast reliability. This application constructs a confidence interval based on the final predicted value of each type of load and the residual standard deviation of that type of load. The confidence interval provides the range within which the predicted value may fall, offering an interval estimate of forecast uncertainty. By comprehensively considering the fluctuation of both the predicted value and the error, it can more comprehensively reflect the reliability of load forecasting. By comparing the lower limit of the confidence interval with the minimum stable operating load threshold, this application can promptly identify potential equipment operation risks caused by load forecasting. When the width of the confidence interval exceeds the uncertainty tolerance threshold, it indicates that the forecast uncertainty is too large, and the actual load may deviate significantly from the predicted value, which may pose a risk to the operation of the energy system. The early warning signal can promptly provide decision-makers with feedback on the uncertainty of load forecasting and equipment operation risks, providing an important basis for adjusting operating strategies.

[0026] Optionally, after issuing the warning signal, the method further includes: Collect multi-source load data and meteorological mutation factors at the current moment, calculate the deviation rate between the multi-source load data at the current moment and the load forecast value at the current moment, and set a dynamic correction threshold based on the meteorological mutation factors. Obtain the meteorological mutation factor at the current moment. If the deviation rate exceeds the dynamic correction threshold corresponding to the meteorological mutation factor at the current moment, use a sliding window to extract the historical multi-source load data of the most recent N time steps. Use the historical multi-source load data of N time steps, the historical meteorological data of the corresponding time steps, and the corresponding coupled feature vector to fine-tune the hidden layer state of the multi-task LSTM model. Use the fine-tuned multi-task LSTM model to re-predict the predicted values ​​of each frequency domain component at the next moment.

[0027] This application collects multi-source load data and meteorological mutation factors at the current moment, and calculates the deviation rate between the multi-source load data at the current moment and the load forecast value at the current moment. It can accurately quantify the degree of difference between the actual load and the forecast load, and clarify the magnitude of the forecast error at the current moment.

[0028] This application sets a dynamic correction threshold based on meteorological abrupt change factors, fully considering the impact of meteorological factors on load forecasting. Different meteorological abrupt changes have varying degrees of impact on load. The dynamic correction threshold can be flexibly adjusted according to the intensity of meteorological abrupt change factors, enabling the model correction mechanism to be triggered in a timely manner when meteorological conditions change significantly, thereby improving the accuracy of forecasts.

[0029] If the deviation rate exceeds the dynamic correction threshold, a sliding window is used to extract historical multi-source load data from the most recent N time steps. This historical data, along with historical meteorological data from the corresponding time steps and the corresponding coupled feature vectors, are then used to fine-tune the hidden layer state of the multi-task LSTM model. The multi-task LSTM model has strong time-series data processing capabilities and can capture the complex relationships between load and meteorological data. By fine-tuning the hidden layer state, the model can quickly adapt to current load change trends and meteorological conditions, thereby improving prediction accuracy. Re-predicting the predicted values ​​of each frequency domain component at the next time step using the fine-tuned model can more accurately reflect load changes in different frequency domains, improving overall prediction accuracy.

[0030] Optionally, when training a multi-task LSTM model using load components, historical meteorological data, and coupled feature vectors in different frequency domains, the method further includes: An attention mechanism is used to weight the load components, historical meteorological data and coupled feature vectors in different frequency domains, calculate the correlation score between each data feature and the current prediction task, and dynamically adjust the weight of each data feature in the model training process based on the correlation score.

[0031] Energy load data typically contains a variety of complex information. The importance of load components in different frequency domains, historical meteorological data, and coupled feature vectors to the prediction task varies. Attention mechanisms can calculate the relevance scores of each data feature to the current prediction task. These scores are used to dynamically adjust the weights of each data feature during model training, allowing the model to focus more on key features highly relevant to the current prediction task and reduce attention to irrelevant or low-relevance features. During the weighting process, the model can learn the association patterns between different features, thereby better understanding how these features collectively influence the load prediction task.

[0032] Secondly, this application provides a multi-source load forecasting system for regional energy systems, which adopts the following technical solution: A regional energy system multi-source load forecasting system includes: a processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method as described in the first aspect.

[0033] In summary, this application includes at least one of the following beneficial technical effects: 1. This application constructs a multi-source load coupling correlation graph using electrical, cooling, and heating loads as nodes and the rated thermodynamic conversion parameters of equipment as edge weights. Topological features are extracted from this graph, and coupling feature vectors are constructed based on these features. This fully utilizes the coupling relationships between loads to improve prediction accuracy. This application also constructs a multi-task LSTM model. By sharing some model parameters, the multi-task LSTM model can utilize the correlations between different types of loads, improving the model's generalization ability and prediction accuracy.

[0034] 2. This application uses load components in different frequency domains, historical meteorological data, and coupled feature vectors to train a multi-task LSTM model. Load components in different frequency domains provide the characteristics of load at different time scales; historical meteorological data takes into account the impact of external environmental factors on load; and coupled feature vectors reflect the interrelationships between different types of loads. By integrating these features into the model, the multi-task LSTM model can learn the load change patterns from multiple perspectives, thereby improving the accuracy of prediction.

[0035] 3. This application reconstructs the predicted values ​​of electricity, cooling, and heating loads by performing maximum overlap discrete wavelet inverse transform on the predicted values ​​of each frequency domain component. This can better handle the complex features in the load data, reduce prediction errors, and improve the accuracy of prediction. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method for decomposing data from S11 to S13 for prediction in Embodiment 1 of this application; Figure 2 This is a flowchart of the method for dynamically adjusting weights from S141 encoding to S143 in other embodiments of this application; Figure 3 This is a flowchart of Embodiment 3 of this application. Detailed Implementation

[0037] The following combination Figures 1 to 3 This application will be described in further detail.

[0038] Example 1: This example discloses a multi-source load forecasting method for regional energy systems, referring to... Figure 1The method includes: S11 data decomposition, S12 constructing coupled feature vectors, and S13 prediction. In this embodiment, historical multi-source load, meteorological data, and equipment parameters of the regional energy system are first obtained. The load data is then decomposed using maximum overlap discrete wavelet transform, and a coupled correlation graph with electricity, cooling, and heating loads as nodes and equipment parameters as weights is constructed to extract topological features. Subsequently, a multi-task LSTM model is constructed, and multi-task learning and training are performed by combining different frequency domain load components, historical meteorological data, and coupled feature vectors. Finally, the meteorological forecast and feature vectors are input to predict each frequency domain component, and the model is reconstructed using inverse wavelet transform to achieve accurate prediction of electricity, cooling, and heating loads. The execution process of each step in this embodiment is as follows: S11 data decomposition obtains historical multi-source load data, historical meteorological data, and rated thermodynamic conversion parameters of equipment in the regional energy system.

[0039] Historical multi-source load data: time series of electrical load, cooling load, and heating load, with sampling intervals uniformly set to a fixed step size (e.g., 15 min, 30 min, 1 h).

[0040] Historical meteorological data: Time series data of exogenous meteorological factors affecting load, such as ambient temperature, relative humidity, solar radiation intensity, and wind speed.

[0041] Rated thermodynamic conversion parameters of equipment: such as the coefficient of performance (COP) of refrigeration units, the coefficient of performance (COP) of heat pump units, and the waste heat recovery efficiency of waste heat recovery devices. These parameters can be obtained from the equipment's instruction manual or related technical documents.

[0042] The historical multi-source load data is decomposed using Maximum Overlap Discrete Wavelet Transform (MODWT) to obtain load components in different frequency domains. The process is as follows: By selecting appropriate wavelet basis functions (such as Daubechies wavelet functions) and decomposition level n, the original load time series is decomposed into sub-components of different frequency domain scales using MODWT. The computational model for decomposition is as follows:

[0043] in, Historical multi-source load data; n is the number of decomposition layers; Let i be the load component of the load data in the i-th frequency domain; The low-frequency component represents the load; among them, the high-frequency component reflects the random fluctuation of the load, the medium-frequency component reflects the intraday periodic variation of the load, and the low-frequency component reflects the long-term trend of the load.

[0044] S12 constructs a coupling feature vector, using electrical load, cooling load, and heating load as nodes, and the rated thermodynamic conversion parameters of the equipment as the initial values ​​of the edge weights, to construct a multi-source load coupling correlation graph.

[0045] In this embodiment, the types and weights of directed edges in the multi-source load coupling correlation graph are as follows: The weight of the directed side electrical load to cooling load is the rated refrigeration coefficient of the refrigeration unit, which represents the electro-refrigeration conversion efficiency; The weight of the directed edge electrical load to heat load is the rated heating coefficient of the heat pump, which represents the electric heating conversion efficiency; The weight of directed edge heat load → cold load is the rated thermodynamic coefficient of the waste heat driven refrigeration system (such as an absorption chiller), which characterizes the rated conversion efficiency of heat energy to cold energy.

[0046] Using relevant algorithms from graph theory, topological features are extracted from a multi-source load coupling graph. These features include node degree, clustering coefficient, and shortest path length. Finally, these topological features are combined into a coupling feature vector according to a preset order. Node degree refers to the number of edges directly connected to the current node. For a node, its clustering coefficient is the ratio of the actual number of edges between it and its neighbors to the maximum possible number of edges, i.e., the local clustering coefficient. The shortest path length represents the minimum sum of weights of the directed edges traversed between two nodes.

[0047] In other embodiments, the method further includes: Identify coupled devices within the regional energy system, including chiller units, heat pump units, and waste heat recovery devices. Construct characteristic curves of the coefficient of performance (COP) of each coupled device as a function of partial load factor (PLR). The calculation model is as follows:

[0048] in, , , The regression coefficients to be fitted; This represents the partial load factor.

[0049] Extract multiple sets of sample data from historical operational data: The sum of squared errors function is constructed, and its calculation model is as follows:

[0050] in, This is the sum of squared errors; Let i be the energy efficiency coefficient in the i-th group of sample data; Let M be the partial load rate in the i-th sample data; M is the number of samples.

[0051] The regression coefficients were obtained by fitting the data using the least squares method, with constraints set during the fitting process: , and .

[0052] Partial load factor (PLR) is the current output power of the equipment. With rated output power The ratio, its calculation model is as follows:

[0053] According to the characteristic curve Establish a mapping relationship between electrical power input and cooling / heating power output, that is:

[0054] in, Current input power; This represents the current output power.

[0055] Under rated operating conditions,

[0056] in, This is the rated input power.

[0057] Substituting the above relationship into the calculation model of the characteristic curve, the calculation model of the mapping relationship is obtained as follows:

[0058] in, This refers to the current output power, i.e., the cooling / heating power output. This refers to the current input power, i.e., the electrical power input. , , The regression coefficients to be fitted; This is the rated output power.

[0059] Numerical differentiation is used to calculate the marginal contribution rate of a unit change in electrical load to changes in cooling and heating loads based on the established mapping relationship. For example, at a specific partial load factor, this results in a small change in electrical power input. Measure the corresponding change in cooling load. and heat load variation The marginal contribution rate of electrical load to cooling load is... The marginal contribution rate to heat load is .

[0060] Based on the calculated marginal contribution rate, the weights of the corresponding edges in the multi-source load coupling correlation graph are updated. For example, if the marginal contribution rate of electrical load to cooling load is different from the rated coefficient of performance of the refrigeration unit, the weights of the edges between electrical and cooling loads are updated to the calculated marginal contribution rate.

[0061] S13 prediction uses a deep learning framework to build a multi-task LSTM model. The network structure of the multi-task LSTM model includes: Input layer: Used to receive input data; Multiple LSTM layers: LSTM layers are used to process sequential data and capture long-term dependencies in the data. In this embodiment, the multi-task LSTM model contains 2-3 LSTM layers, each containing 64 neurons.

[0062] Fully connected layer: Used to map the output of the last LSTM layer to the predicted values ​​of electrical, cold, and hot load components in different frequency domains. The number of neurons in the fully connected layer is determined according to the number of prediction tasks. In this embodiment, each electrical, cold, and hot load component in each frequency domain corresponds to one output branch. In this embodiment, the multi-source load data is decomposed into 3 frequency domains, so the number of neurons in the fully connected layer is 9 (3 frequency domains × 3 load types).

[0063] The multi-task LSTM model is trained using load components, historical meteorological data, and coupled feature vectors in different frequency domains. During the training process, the load components, historical meteorological data, and coupled feature vectors in different frequency domains are used as input data, and the corresponding actual electricity, cooling, and heating load component data in different frequency domains are used as output data to train the constructed multi-task LSTM model.

[0064] This embodiment uses mean squared error (MSE) as the loss function to measure the difference between the predicted and actual values; it uses the Adam optimization algorithm to update the model parameters to minimize the loss function; the number of training epochs is 100-200; and the batch size is set to 32-128.

[0065] The historical multi-source load data with a preset duration between the current time and the target multi-source load data is recorded as the target multi-source load data. The target multi-source load data is decomposed by maximum overlap discrete wavelet transform (MODWT) to obtain the first target load component.

[0066] The first target load component, meteorological forecast data, and coupled feature vector are input into the trained multi-task LSTM model to obtain the predicted values ​​of each frequency domain component. The predicted values ​​of each frequency domain component are then reconstructed by the maximum overlap discrete wavelet inverse transform to obtain the predicted values ​​of electricity, cooling, and heating loads.

[0067] Reconstructing the predicted values ​​of each frequency domain component by performing maximum overlap discrete wavelet inverse transform means linearly superimposing the predicted values ​​of each frequency domain component.

[0068]

[0069] in This represents the predicted load data, including predicted values ​​for electricity, cooling, and heating loads; n represents the number of decomposition levels. This is the predicted value of the load component in the i-th frequency domain of the load data; These are the predicted values ​​for the low-frequency components.

[0070] Reference Figure 2 In other embodiments, when training a multi-task LSTM model using load components, historical meteorological data, and coupled feature vectors in different frequency domains, the method further includes: S141 encoding, S142 calculating correlation scores, and S143 dynamically adjusting weights.

[0071] S141 coding involves standardizing and uniformly encoding load components, historical meteorological data, and coupled feature vectors in different frequency domains. The process is as follows: Each frequency domain load component obtained from the S11 data decomposition is time-series encoded, with each frequency domain component corresponding to a feature vector. The encoding dimension is... ,in, For the feature dimension of a single frequency domain component, This represents the timing length.

[0072] After normalizing historical meteorological data, it is converted into feature vectors using a time-series coding method, with a coding dimension of [missing value]. ,in, As a dimension of meteorological characteristics, The timing length is consistent with that of the frequency domain components.

[0073] The coupled feature vectors are encoded, and considering the time-varying nature of the coupled features, they are transformed into time-series feature vectors with an encoding dimension of [dimension missing]. ,in, For the dimension of the coupling feature, The timing length is consistent with that of the frequency domain components.

[0074] The three types of encoded feature vectors are concatenated to obtain a unified input feature matrix with dimension . .

[0075] S142 calculates the relevance score, defining the query vector Q, key vector K, and value vector V for the attention mechanism. The concatenated input feature matrix is ​​mapped to the query vector Q, key vector K, and value vector V, respectively. Here, the query vector Q corresponds to the required features of the current prediction task, obtained by mapping from the hidden layer output of the multi-task LSTM model, and has a dimension of [missing information]. The key vector K is obtained by mapping from the input feature matrix, and its dimension is... The feature attributes used to characterize each input feature; the value vector V is directly derived from the input feature matrix, with dimension V. , which are the original feature values ​​used to characterize each input feature.

[0076] The relevance score between query vector Q and key vector K is calculated using the additive attention formula. The formula is as follows:

[0077] in, , , These are trainable weight matrices for the query vector, key vector, and value vector, respectively, which are adaptively updated through model training. The function is used for non-linear mapping, which maps the correlation score to the interval [-1, 1].

[0078] The relevance scores are normalized using the Softmax function to obtain the attention weights for each input feature. The formula is as follows:

[0079] Normalized attention weights The dimension is the same as the input feature matrix, and each element The weight of the i-th feature at the j-th time step represents the strength of the feature's relevance to the prediction task at that time step.

[0080] S143 dynamically adjusts the weights, using the calculated attention weights. The input feature matrix after attention enhancement is obtained by weighted summation with the value vector V, as shown in the following formula:

[0081] in, The input feature matrix after attention enhancement; Attention weights; V is the transposed value vector.

[0082] The dimension of the attention-enhanced input feature matrix is ​​the same as that of the original input feature matrix. However, the value of each feature in the attention-enhanced input feature matrix is ​​adjusted according to the relevance of the current prediction task. That is, features with high relevance scores have greater attention weights and a higher proportion in the attention-enhanced input feature matrix, while features with low relevance scores have smaller attention weights and a lower proportion in the attention-enhanced input feature matrix.

[0083] The input feature matrix after attention enhancement Instead of the traditional equal-weighted input, a multi-task LSTM model is input for training. During backpropagation, the multi-task LSTM model not only updates its own network parameters but also simultaneously updates the attention mechanism. , , These are trainable weight matrices for the query vector, key vector, and value vector, respectively, which make the calculation of attention weights more aligned with the needs of the prediction task, thereby enabling adaptive adjustment of feature weights.

[0084] This embodiment uses an attention mechanism to weight different frequency domain load components, historical meteorological data, and coupled feature vectors to achieve dynamic optimization of input features. This enables the multi-task LSTM model to more accurately capture the correlation between multi-source features and the load forecasting task, further improving the accuracy of the multi-task LSTM model's prediction results.

[0085] Example 2: This example differs from Example 1 in that the method further includes: The system collects the operating parameters of the coupled equipment in real time, including: partial load rate, actual input power, and actual output cooling / heating power. Based on the real-time collected input and output power, the system calculates the actual energy efficiency ratio of the equipment at the current moment.

[0086] Based on the actual partial load factor at the current moment, substitute it into the characteristic curve model to obtain the theoretical energy efficiency ratio at the current moment.

[0087] For each coupled device, the deviation between its actual energy efficiency ratio and theoretical energy efficiency ratio is calculated in real time. The calculation model for the deviation is as follows:

[0088] in, Let t be the deviation of the energy efficiency ratio of the coupling device at time t; This refers to the actual energy efficiency ratio; This is the theoretical energy efficiency ratio.

[0089] When the deviation continues to exceed the allowable error range (0.05~0.1 in this embodiment) within a preset time window (e.g., 4 sampling periods), it indicates that the coupling device is operating abnormally, and its actual energy conversion efficiency can no longer be characterized by the characteristic curve. The marginal contribution rate calculated based on the device is no longer of reference value. At this time, it is necessary to make an emergency adjustment to the weight of the directed edge corresponding to the device in the multi-source load coupling correlation diagram. That is, the weight of the corresponding edge of the coupling device in the multi-source load coupling correlation diagram is reset to zero until the deviation does not exceed the allowable error range. Then, the marginal contribution rate is recalculated, and the weight of the corresponding edge in the multi-source load coupling correlation diagram is updated using the recalculated marginal contribution rate.

[0090] In other embodiments, the deviation between the actual energy efficiency ratio and the theoretical energy efficiency ratio of each coupling device can also be equal to the absolute difference between the actual energy efficiency ratio and the theoretical energy efficiency ratio. In this embodiment, the allowable error range is 0.1 to 0.5.

[0091] In other embodiments, the method further includes: Based on the load characteristics of the regional energy system, the event types and corresponding event tags corresponding to historical multi-source load data are identified, and an event database containing event tags for statutory holidays, exhibitions, and events is established. The event database includes: event tags (unique identifiers); event time range (start time, end time), accurate to the time granularity consistent with load forecasting and data collection (e.g., hour, day); event impact range (the coverage area of ​​the corresponding regional energy system, identifying the load nodes affected by the event); and event load impact characteristics (e.g., sudden increase / decrease in load, type of load affected (electricity / cooling / heating), maximum load offset).

[0092] For each event tag, a corresponding load offset correction coefficient matrix is ​​constructed. The load offset correction coefficient matrix has a dimension of m×1, where m is the dimension of the coupled feature vector. Each element in the load offset correction coefficient matrix is ​​calculated based on the historical load disturbance data corresponding to the event tag. Specifically, the historical multi-source load data and topology features of the event tag are extracted, and the ratio of the actual value of the topology feature at the time of the event to the baseline value of the topology feature under normal conditions (no event) is calculated. The average value of the ratios of multiple event tags of the same type is calculated as the correction coefficient corresponding to the topology feature.

[0093] For example, a record in the event database has the following event tags: statutory holiday: Spring Festival; time range: January 29 to February 4; event impact range: residential area energy system and industrial area energy system; event load impact characteristics: residential area electricity load and heat load increase sharply, industrial area cooling load and electricity load decrease sharply, and the historical maximum load offset is 15%.

[0094] Using a regular week (without holidays) as the baseline, calculate the average topological eigenvalue: ; Obtain the topological feature values ​​at the same time during the Spring Festival in the past three years: , , .

[0095] The correction factor for node degree is: (4.5 + 4.2 + 4.4) ÷ 6.0 ÷ 3 ≈ 0.728; The correction factor for the clustering coefficient is: (0.5 + 0.45 + 0.48) ÷ 0.8 ÷ 3 ≈ 0.596; The path length correction factor is: (6.67+7.14+6.9)÷4.2÷3≈1.64; The correction coefficient matrix is .

[0096] If the prediction period contains any event label, then the Hadamard product operation is performed between the correction coefficient matrix corresponding to the prediction period and the topological features to generate a coupled feature vector containing the event perturbation factor.

[0097] Example 3: Reference Figure 3 The difference between this embodiment and Embodiment 1 is that, after performing the prediction in S13, the method further includes: S31 residual correction obtains the actual measured values ​​of multi-source loads and calculates the residual sequence between the model prediction value and the actual measured value at the corresponding time.

[0098] The calculation model for the residual value of electrical load is as follows:

[0099] in, This represents the residual value of the electrical load. The actual measured value of the electrical load at time t; This is the model prediction value of the electrical load at time t, which is the predicted value of the electrical load at time t in the S13 prediction.

[0100] The calculation model for the residual value of cooling load is as follows:

[0101] in, This represents the residual value of the cooling load; This is the actual measured value of the cooling load at that moment; The model prediction of the cooling load at the same time as the actual measured value.

[0102] The calculation model for the residual value of heat load is as follows:

[0103] in, This represents the residual value of the heat load; This is the actual measured value of the heat load at that moment; The model prediction of the heat load at the same time as the actual measured value.

[0104] The sign of various residual values ​​indicates the direction of the prediction deviation. A positive residual value indicates that the model prediction value is less than the actual measurement value; a negative residual value indicates that the model prediction value is greater than the actual measurement value; the absolute value of the residual value indicates the magnitude of the deviation.

[0105] The residual values ​​k time points before time t are sorted in chronological order to obtain the residual sequence, i.e., the electrical load residual sequence. Cooling load residual sequence and heat load residual sequence .

[0106] The residual sequence is smoothed using a moving average algorithm to predict the residual value at the next time step, based on the electrical load residual sequence. For example, the process includes the following steps: Set the sliding window size to k, and select the residual sequence within the sliding window with the current time as the endpoint. The arithmetic mean of the residuals within the sliding window is calculated and used as the smoothed residual at the current time step. This smoothed residual is then used as the residual value at the next time step. The calculation model for the smoothed residual at the current time step is as follows:

[0107] in, This represents the residual value of the smoothed electrical load at the current moment; is the residual value of the power-down load at the i-th time within the sliding window; k is the size of the sliding window.

[0108] The predicted values ​​of electricity, cooling, and heating loads for the next time step in S13 are added to the residual values ​​for the next time step to obtain the final predicted values ​​of electricity, cooling, and heating loads for the next time step.

[0109] S32 calculates the interval width and, based on the residual sequence, calculates the residual standard deviation for each type of load, including the residual standard deviation for electrical loads. The calculation model is as follows:

[0110] Where k is the size of the sliding window; This represents the residual value of the power-down load at the i-th time point within the sliding window; This represents the residual value of the smoothed electrical load at the current moment.

[0111] Residual standard deviation of cooling load The calculation model is as follows:

[0112] Where k is the size of the sliding window; This represents the residual value of the cooling load at the i-th time point within the sliding window; This represents the residual value of the cooling load after smoothing at the current moment.

[0113] residual standard deviation of heat load The calculation model is as follows:

[0114] Where k is the size of the sliding window; This represents the residual value of the heat load at the i-th time point within the sliding window; This represents the residual value of the heat load after smoothing at the current moment.

[0115] Confidence intervals for each load type are constructed based on the final predicted value and the standard deviation of the residuals for that load type. The width of the confidence interval is then calculated. In this embodiment, a commonly used 90% confidence level is used, with 1.645 being the Z-value corresponding to the 90% confidence level. The interval width is the upper limit minus the lower limit of the confidence interval. Therefore, in this embodiment, the interval width of the confidence interval for each load type is 3.29 times the standard deviation. The specific calculation process is as follows: The confidence interval for electrical load is ,in, The final predicted value of the electrical load is given, and the width of the confidence interval for the electrical load is... .

[0116] The confidence interval for cooling load is ,in, The final predicted value of the cooling load is given, and the width of the confidence interval for the cooling load is... .

[0117] The confidence interval for heat load is ,in, The final predicted value of the heat load is given, and the width of the confidence interval for the heat load is... .

[0118] S33 alarm: Obtain the minimum stable operating load threshold of energy equipment in the area (i.e., the lowest load value at which the equipment can operate normally and stably, for example, the minimum stable operating load threshold of a 500RT (refrigeration ton) centrifugal chiller is 527.5kW, the minimum stable operating load threshold of an 800kW air source heat pump is 200kW, and the minimum stable operating load threshold of a waste heat recovery device with a rated capacity of 800kW is 240kW) and the preset uncertainty tolerance threshold (i.e., the upper limit of predictive uncertainty that the system can withstand, such as 10%). The minimum stable operating load threshold of the energy equipment is derived from the equipment's factory parameters and historical operating data.

[0119] An early warning signal is issued when the range width of any load exceeds the preset uncertainty tolerance threshold, or when the lower limit of the confidence range of any load is lower than the minimum stable operating load threshold.

[0120] S34 calculates the deviation rate by collecting multi-source load data and meteorological mutation factors at the current moment. The meteorological mutation factors are obtained by collecting meteorological mutation data at the current moment (time t), including instantaneous strong winds, rainstorms, extreme temperature mutations, etc., and using 0-1 encoding (1 indicates the existence of meteorological mutations, and 0 indicates the absence of meteorological mutations).

[0121] A dynamic correction threshold is set based on the meteorological mutation factor. When the meteorological mutation factor is 0, it indicates that there is no meteorological mutation. The dynamic correction threshold is set to 5%~8%. At this time, the load fluctuation is relatively stable, and the deviation rate judgment can be appropriately relaxed. When the meteorological mutation factor is 1, it indicates that there is a meteorological mutation. Meteorological mutations can easily lead to sudden changes in load. Deviations need to be identified and corrected in time. Therefore, the dynamic correction threshold is set to 3%~5% to improve the sensitivity of deviation rate judgment.

[0122] Calculate the deviation rate between the current multi-source load data and the load forecast value at the previous time, whereby the electrical load deviation rate is... The calculation model is as follows:

[0123] in, The actual measured value of the electrical load at time t; Let t be the predicted value of the electrical load at time t.

[0124] The deviation rate of the cooling load The calculation model is as follows:

[0125] in, The actual measured value of the cooling load at time t; Let t be the predicted value of the cooling load at time t.

[0126] The deviation rate of the heat load The calculation model is as follows:

[0127] in, The actual measured value of the heat load at time t; Let t be the predicted value of the heat load at time t.

[0128] All of the above deviation rates are non-negative. The larger the deviation rate, the greater the prediction deviation at the current moment.

[0129] S35 fine-tuning: Obtain the meteorological mutation factor at the current moment. If the deviation rate exceeds the dynamic correction threshold corresponding to the meteorological mutation factor at the current moment, use a sliding window to extract the historical multi-source load data of the most recent N time steps, and extract the current hidden layer state of the multi-task LSTM model, including the short-term memory state and the long-term memory state, as the initial state for fine-tuning.

[0130] The historical multi-source load data at N time steps are decomposed using maximum overlap discrete wavelet transform to obtain load components in different frequency domains, denoted as the second target load component. The second target load component, the historical meteorological data at the corresponding time step, and the corresponding coupled feature vectors are used to fine-tune the multi-task LSTM model. During fine-tuning, this embodiment employs mini-batch gradient descent, updating only the hidden layer states of the multi-task LSTM model to obtain the fine-tuned model. Then, the fine-tuned multi-task LSTM model is used to re-predict the predicted values ​​of each frequency domain component at the next time step.

[0131] If the deviation rate does not exceed the dynamic correction threshold corresponding to the meteorological mutation factor at the current moment, no action will be taken.

[0132] Example 4: This example discloses a multi-source load forecasting system for a regional energy system. The system includes: a processor and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method.

[0133] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for predicting multi-source loads in a regional energy system, characterized in that, include: Historical multi-source load data, historical meteorological data and equipment rated thermodynamic conversion parameters of the regional energy system are obtained. The historical multi-source load data are decomposed by maximum overlap discrete wavelet transform to obtain load components in different frequency domains. Using electrical, cooling, and heating loads as nodes and the rated thermodynamic conversion parameters of the equipment as edge weights, a multi-source load coupling correlation graph is constructed. Topological features are extracted from the multi-source load coupling correlation graph, and coupling feature vectors are constructed based on the topological features. A multi-task LSTM model is constructed and trained using load components, historical meteorological data, and coupled feature vectors in different frequency domains. The load components, meteorological forecast data, and coupled feature vectors corresponding to historical multi-source load data of a preset duration between the current time are input into the trained multi-task LSTM model to obtain the predicted values ​​of each frequency domain component. The predicted values ​​of each frequency domain component are reconstructed by maximum overlap discrete wavelet inverse transform to obtain the predicted values ​​of electricity, cooling, and heating loads.

2. The method for multi-source load forecasting of regional energy systems according to claim 1, characterized in that, The method further includes: Identify the coupled devices within the regional energy system, including chiller units, heat pump units, and waste heat recovery devices. Construct characteristic curves of the energy efficiency ratio of each coupled device as a function of partial load rate. Establish a mapping relationship between electrical power input and cooling / heating power output based on the characteristic curves. Calculate the marginal contribution rate of a unit change in electrical load to the change in cooling and heating load based on the mapping relationship. Update the weights of corresponding edges in the multi-source load coupling correlation graph using the marginal contribution rate.

3. The method for multi-source load forecasting of regional energy systems according to claim 2, characterized in that, The method further includes: calculating in real time the deviation between the actual energy efficiency ratio of the coupling device and the theoretical energy efficiency ratio obtained based on the characteristic curve; when the deviation continues to exceed the allowable error range within a preset time window, the weight of the corresponding edge of the coupling device in the multi-source load coupling association diagram is reset to zero until the deviation does not exceed the allowable error range.

4. The method for multi-source load forecasting of regional energy systems according to claim 2, characterized in that, The method further includes: Establish an event library containing event tags for statutory holidays, exhibitions, and sporting events; and construct a corresponding load offset correction coefficient matrix for each event tag. If the prediction period matches any event label, then the Hadamard product operation is performed between the correction coefficient matrix corresponding to the prediction period and the topological features to generate a coupled feature vector containing the event perturbation factor.

5. The method for multi-source load forecasting of regional energy systems according to any one of claims 1-4, characterized in that, The method further includes: Obtain the actual measured values ​​of multi-source loads, calculate the residual sequence between the predicted value and the actual measured value at the corresponding time, smooth the residual sequence using the moving average algorithm, predict the residual value at the next time, add the predicted value of the electricity, cooling and heating load at the next time to the residual value at the next time, and obtain the final predicted value of the electricity, cooling and heating load at the next time.

6. The method for multi-source load forecasting of regional energy systems according to claim 5, characterized in that, The method further includes: The residual standard deviation of each type of load is calculated based on the residual sequence. The confidence interval of each type of load is constructed based on the final predicted value of each type of load and the residual standard deviation of that type of load. The width of the confidence interval is then calculated. The system acquires the minimum stable operating load threshold and the preset uncertainty tolerance threshold for energy equipment within the region. When the interval width exceeds the uncertainty tolerance threshold, or the lower limit of the confidence interval is lower than the minimum stable operating load threshold, an early warning signal is issued.

7. The method for multi-source load forecasting of regional energy systems according to claim 6, characterized in that, After issuing the warning signal, the method further includes: Collect multi-source load data and meteorological mutation factors at the current moment, calculate the deviation rate between the multi-source load data at the current moment and the load forecast value at the current moment, and set a dynamic correction threshold based on the meteorological mutation factors. Obtain the meteorological mutation factor at the current moment. If the deviation rate exceeds the dynamic correction threshold corresponding to the meteorological mutation factor at the current moment, use a sliding window to extract the historical multi-source load data of the most recent N time steps. Use the historical multi-source load data of N time steps, the historical meteorological data of the corresponding time steps, and the corresponding coupled feature vector to fine-tune the hidden layer state of the multi-task LSTM model. Use the fine-tuned multi-task LSTM model to re-predict the predicted values ​​of each frequency domain component at the next moment.

8. The method for multi-source load forecasting of regional energy systems according to any one of claims 1-4, characterized in that, When training a multi-task LSTM model using load components, historical meteorological data, and coupled feature vectors in different frequency domains, the method further includes: An attention mechanism is used to weight the load components, historical meteorological data and coupled feature vectors in different frequency domains, calculate the correlation score between each data feature and the current prediction task, and dynamically adjust the weight of each data feature in the model training process based on the correlation score.

9. A multi-source load forecasting system for a regional energy system, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method as described in any one of claims 1-8.

Images