Information processing method for integrated energy system, computer device and medium

Abstract

This application discloses an information processing method, computer equipment, and medium for integrated energy systems, belonging to the field of artificial intelligence technology. The method includes the following steps: responding to the input of a training sample set, inputting the training sample set into a multi-task prediction model to obtain a first predicted value for electrical load, a second predicted value for thermal load, a third predicted value for cooling load, and a fourth predicted value for carbon emissions; wherein the training sample set is constructed based on historical data of electrical load, thermal load, cooling load, carbon emissions, and meteorological data; constructing a carbon balance physical constraint loss term; wherein the carbon balance physical constraint loss term is used to measure the deviation between the fourth predicted value and the estimated value of carbon emissions, the estimated value being obtained through a linear coupling relationship; weighted summing of the carbon balance physical constraint loss term and the data-driven loss term to construct a total loss function, and updating the parameters of the multi-task prediction model through backpropagation. This application improves the accuracy of the prediction results.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to an information processing method, computer equipment, and medium for integrated energy systems. Background Technology

[0002] An Integrated Energy System (IES) is a new type of energy system that enables the coordinated coupling and optimized allocation of multiple energy forms, such as electricity, heat, cooling, and renewable energy, across different stages including source, grid, load, and storage. Multi-energy load forecasting is a key foundation for the planning and operational optimization of integrated energy systems. Accurate forecasting of multiple energy variables, such as electrical load, heat load, cooling load, and carbon emissions, provides decision support for system scheduling optimization, capacity allocation, and carbon emission control.

[0003] However, there are complex physical coupling relationships between different energy subsystems in an integrated energy system. Under conditions of insufficient samples, changes in operating conditions, or external disturbances, prediction results that are inconsistent with actual physical laws are easily generated. Summary of the Invention

[0004] The purpose of this application is to provide an information processing method, computer equipment, and medium for integrated energy systems, which can solve the problem of how to improve the physical rationality of prediction results, thereby improving the accuracy of prediction results.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide an information processing method for an integrated energy system, the method comprising: In response to the input of the training sample set, the training sample set is input into the multi-task prediction model to obtain a first predicted value of electrical load, a second predicted value of thermal load, a third predicted value of cooling load, and a fourth predicted value of carbon emissions; wherein, the training sample set is constructed based on historical data of electrical load, thermal load, cooling load, carbon emissions, and meteorology. A carbon balance physical constraint loss term is constructed; wherein, the carbon balance physical constraint loss term is used to measure the deviation between the fourth predicted value and the estimated value of carbon emissions, and the estimated value is calculated by the first predicted value, the second predicted value and the third predicted value through a linear coupling relationship; The total loss function is constructed by weighted summing of the carbon balance physical constraint loss term and the data-driven loss term, and then updated by backpropagation of the parameters of the multi-task prediction model.

[0006] In a second aspect, embodiments of this application provide a computer device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0007] Thirdly, embodiments of this application provide a computer-readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0008] Fourthly, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0009] In this embodiment, a carbon balance physical constraint loss term is constructed, and a weighted sum of the carbon balance physical constraint loss term and the data-driven loss term is performed to construct a total loss function. The parameters of the multi-task prediction model are then updated through backpropagation. This achieves the embedding of carbon emission mechanism constraints into the model training process, i.e., a deep integration of data-driven and physical constraints. This ensures that the trained multi-task prediction model is simultaneously subject to data-driven learning and physical constraints, improving the physical rationality of the model's prediction results and thus enhancing the accuracy of the predictions. Attached Figure Description

[0010] Figure 1 This is one of the flowcharts illustrating an information processing method for an integrated energy system provided in some embodiments of this application; Figure 2 These are schematic diagrams of the improved WConv structure provided by some embodiments of this application; Figure 3 These are schematic diagrams of the improved WCKAN model provided in some embodiments of this application; Figure 4 These are schematic diagrams illustrating the structure of the Informer as described in some embodiments of this application; Figure 5 This is one of the flowcharts illustrating an information processing method for an integrated energy system provided in some embodiments of this application; Figure 6 These are internal structural diagrams of a computer device provided in some embodiments of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0013] In one exemplary embodiment, this application proposes an information processing method for an integrated energy system. The information processing method for an integrated energy system provided by this application will be described in detail below with reference to the accompanying drawings and specific embodiments and application scenarios.

[0014] Reference Figure 1 The method includes steps 102-106. Wherein: Step 102: In response to the input of the training sample set, the training sample set is input into the multi-task prediction model to obtain the first predicted value of electrical load, the second predicted value of thermal load, the third predicted value of cooling load, and the fourth predicted value of carbon emissions; wherein, the training sample set is constructed based on historical data of electrical load, thermal load, cooling load, carbon emissions, and meteorology.

[0015] In some embodiments, firstly, historical data of the integrated energy system are collected, including historical meteorological data, historical electricity load data, historical heat load data, historical cooling load data, and historical carbon emission data. This historical data is then organized and spliced ​​in chronological order to construct a multivariate time series dataset. Secondly, the multivariate time series dataset is preprocessed to select highly correlated variables and appropriate sliding time windows, dividing the multivariate time series dataset into input samples and corresponding label data to form a training sample set.

[0016] In some embodiments, preprocessing includes: missing value imputation, normalization, and autocorrelation and cross-correlation analysis of variables in historical data.

[0017] In some embodiments, carbon emissions, electricity load, cooling load, and heating load data of the integrated energy system at Arizona State University (ASU) Tempe Campus can be collected from January to December 2020, and from February to June 2021, with a time resolution of 1 hour. Meteorological data for the corresponding time periods is also acquired, sourced from the National Solar Radiation Database (NSRDB), including meteorological characteristics such as temperature, precipitation, and air pressure. Furthermore, timestamp information, holiday information, and weekday information can be used as input feature variables.

[0018] In some embodiments, during the data acquisition process, some data may be missing or abnormal due to environmental factors, hardware failures, manual input errors, or abnormal data transmission. For missing and outlier values, deletion or linear interpolation methods can be used to clean the original historical data to obtain a complete and reliable data sequence.

[0019] In some embodiments, to eliminate the dimensional differences between different feature variables and improve the stability of model training, the cleaned data is normalized. For example, the Min-Max normalization method is used to map each feature data to the interval [0,1].

[0020] In some embodiments, through the above correlation analysis, electrical load, heat load, cooling load, carbon emissions, temperature, precipitation, and atmospheric pressure are ultimately selected as input feature variables for the model.

[0021] In some embodiments, to determine the window length of the sliding time window for time series modeling, autocorrelation function (ACF) analysis can be performed on the load series. For example, the window length can be determined by autocorrelation analysis and set to 24 hours. This window length can effectively characterize the diurnal cyclical variation characteristics and short-term time series dependencies of the integrated energy system load and carbon emission data, while avoiding the introduction of too much redundant historical information, thereby reducing the computational complexity of the model while ensuring prediction accuracy.

[0022] In some embodiments, the multi-task prediction model may include, but is not limited to, a deep learning prediction model built on recurrent neural networks, Transformer models, and combinations thereof.

[0023] Step 104: Construct a carbon balance physical constraint loss term; wherein the carbon balance physical constraint loss term is used to measure the deviation between the fourth predicted value and the estimated value of carbon emissions, and the estimated value is calculated by the first predicted value, the second predicted value and the third predicted value through a linear coupling relationship.

[0024] In some embodiments, the linear coupling relationship specifically refers to: The calculated value is equal to the sum of the learnable error compensation term and the first product, the second product, and the third product; Wherein, the first product is the product of the first predicted value and the learnable electrical carbon emission coefficient, the second product is the product of the second predicted value and the learnable thermal carbon emission coefficient, and the third product is the product of the third predicted value and the learnable cold carbon emission coefficient; the learnable error compensation term and the learnable carbon emission coefficient are jointly optimized with the parameters of the multi-task prediction model through backpropagation during the training process.

[0025] Specifically, this embodiment introduces Physical-Informed Neural Network (PINN) constraints to embed the carbon emission mechanism in the integrated energy system into the model training process. In the integrated energy system, the total carbon emissions... It can be considered as being caused by electrical load Cooling load and heating load The result of joint contributions. Based on the assumption of linear superposition of energy conservation and emission factors, an approximate linear coupling relationship can be established as shown in Equation 1 below.

[0026] (1) in, It is a learnable error compensation term used to compensate for baseline emissions, unmodeled stationary emissions, or systematic errors; These are the carbon emission coefficients corresponding to electricity, cooling, and heat, namely the electricity carbon emission coefficient, the heat carbon emission coefficient, and the cooling carbon emission coefficient. They can be understood as "the increase in carbon emissions caused by a unit of electricity load, cooling load, and heating load", and their unit is tons.

[0027] Correspondingly, the physical residual, or deviation, can be represented by the following formula 2.

[0028] (2) in, , , , These are the predicted electricity, cooling, and heating loads, as well as carbon emissions, namely the first, second, third, and fourth predicted values. This is a calculated value.

[0029] In some embodiments, the carbon balance physical constraint loss term can be calculated using the weighted mean square error form, and its expression is shown in Formula 3 below.

[0030] (3) in, The total number of samples, For sample index.

[0031] Step 106: The carbon balance physical constraint loss term and the data-driven loss term are weighted and summed to construct the total loss function, and the parameters of the multi-task prediction model are updated through backpropagation.

[0032] In some embodiments, the total loss function of the PINN-based multi-task prediction model is obtained by jointly optimizing the data-driven loss term and the carbon balance physical constraint loss term, as shown in Equation 4.

[0033] (4) in, This is the second weighting coefficient, used to balance the data fitting accuracy and physical consistency constraints. This is the data-driven loss term, i.e., the loss function of the multi-task prediction model without PINN. This represents the set of trainable parameters for the model.

[0034] This embodiment constructs a carbon balance physical constraint loss term and weights it together with the data-driven loss term to build a total loss function. The parameters of the multi-task prediction model are then updated through backpropagation. This embeds carbon emission mechanism constraints into the model training process, achieving a deep integration of data-driven and physical constraints. The trained multi-task prediction model is simultaneously subject to data-driven learning and physical constraints, improving the physical plausibility of the model's predictions and thus enhancing their accuracy.

[0035] It should be noted that the following embodiments are for the case where the multi-task prediction model includes weighted convolution and the Chebyshev-Kolmogorov-Arnold Network (WCKAN) module, and are various optional improvements to the WCKAN module.

[0036] In some embodiments, the WCKAN module includes a Chebyshev Kolmogorov–Arnold Network (ChebyshevKAN) branch based on Chebyshev multinomials, used for feature fusion of the input tensor to obtain spatial coupling features; wherein, the spatial coupling features are used to describe the nonlinear spatial coupling relationship between different feature variables in the input tensor; the process of fusing to obtain the spatial coupling features is as follows: The input tensor to the ChebyshevKAN branch is normalized to the interval [-1, 1] using the hyperbolic tangent function to ensure that the tensor satisfies the domain requirement of the Chebyshev polynomial. For each feature dimension of the normalized input tensor, a Chebyshev polynomial basis function of order m is constructed, where m is a configurable hyperparameter. Learnable coefficients are assigned to the function values ​​of each order of the Chebyshev polynomial, and the function values ​​of each order are weighted and summed to obtain the feature representation corresponding to that dimension. The feature representations of all dimensions of the normalized input tensor are concatenated into vectors to obtain the spatial coupling features.

[0037] The input tensor is obtained by preprocessing the training sample set.

[0038] In some embodiments, the value of each Chebyshev polynomial is obtained by performing an inverse cosine transform on the characteristic variable and then taking its cosine.

[0039] This embodiment constructs a ChebyshevKAN branch to achieve global nonlinear mapping and fusion of multivariate features, effectively characterizing the high-order nonlinear spatial coupling relationship between electrical, thermal, cooling loads and carbon emissions, and improving the ability to model multiple energy variables.

[0040] In some embodiments, refer to Figure 2 The WCKAN module further includes a weighted convolution (WConv) branch, used to extract the temporal features of each variable in the input tensor; wherein, the input tensor is obtained by preprocessing the training sample set; the process of extracting the temporal features is as follows: The weighted convolution branch sets a set of learnable first weight coefficients for each feature channel. The first weight coefficients are constructed based on the distance between each position of the convolution kernel and the current time step, and decay exponentially as the distance increases. The first weight coefficients are normalized. The normalized first weight coefficients are multiplied element-wise with the standard convolution kernel parameters to obtain a learnable weighted convolution kernel. The weighted convolution kernel is slid along the time dimension with a stride of 1 to perform convolution operations to obtain the time features.

[0041] It should be noted that, in order to make the convolution operator (i.e., the weighted convolution branch) more effectively applied to one-dimensional time series signals, this embodiment improves the traditional convolution structure by introducing a set of adaptive weight coefficients that vary with time distance before the convolution kernel operation.

[0042] Because time series data typically exhibit significant local correlations, data from adjacent time steps usually contribute more significantly to the features of the current moment, while historical signals from farther away have a relatively weaker impact on current predictions. Therefore, by setting a first weight coefficient that decays exponentially with time distance, the modeling ability of the convolution operator for time features can be improved.

[0043] The following uses historical data of electrical load as an example to illustrate the process of extracting time features. Examples of other historical data in the training sample set will not be repeated here.

[0044] With a convolution kernel length of In this case, an adaptive first weighting coefficient corresponding to the electrical load is introduced. The calculation formula is as follows: in, Represents the first in the convolution kernel One location, Indicates the kernel length. ; Indicates the first The first weighting coefficient corresponding to the electrical load at each location.

[0045] In some embodiments, in order to ensure the stability of the weights at each time step and to avoid the convolution result being offset by weights that are too large or too small, the first weight coefficients are normalized so that the sum of all weight coefficients is 1.

[0046] In some embodiments, the convolution kernel parameters (i.e. standard convolution kernel parameters) and the corresponding weight coefficient matrix of the electrical load feature channels are constructed.

[0047] Let the convolution kernel parameters of the electrical load characteristic channel be... and the corresponding first weight coefficient matrix They are respectively: By performing element-wise multiplication of the first weight coefficient with the convolution kernel parameters, the weighted convolution kernel parameters can be obtained, and their expression is as follows: in, This represents the weighted convolution kernel parameters; Indicates the convolution kernel parameters; This represents the corresponding first weight coefficient.

[0048] At time step When performing convolution calculations on the electrical load sequence in the input tensor, the convolution output can be expressed as: in, Represents electrical load sequence In time The value at that location; Indicates the electrical load sequence at time step The feature output obtained after convolution calculation.

[0049] Furthermore, the outputs of each feature channel are concatenated along the channel dimension to obtain the WConv temporal feature representation.

[0050] It should be noted that when When all values ​​are 1, the weighted convolution operation described above degenerates into a traditional one-dimensional convolution operation. By introducing time decay weights, the model's focus on recent time information can be enhanced, thereby improving its ability to extract time features from electrical load time series and further improving the expressive power of the prediction model.

[0051] This embodiment enhances the ability of the multi-task prediction model to capture local dynamic changes in the time dimension by introducing weighted convolution with a learnable exponential decay mechanism, and can more accurately reflect the short-term fluctuation characteristics of the load.

[0052] In some embodiments, see Figure 3 The WCKAN module further includes a gated residual fusion structure, used to adaptively fuse the spatial coupling features and the temporal features to obtain spatiotemporal fusion features; the process of fusing to obtain the spatiotemporal fusion features is as follows: Initialize the learnable gating parameters.

[0053] The learnable gating parameters are normalized using the Sigmoid activation function to obtain the gating coefficients.

[0054] The spatial coupling feature and the temporal feature are added element by element, and then multiplied by the gating coefficient to obtain the weighted fusion feature.

[0055] The weighted fusion feature is added element-wise to the original feature in the tensor to obtain the spatiotemporal fusion feature.

[0056] In some embodiments, the ChebyshevKAN branch mainly utilizes the high-order approximation capability of Chebyshev polynomials to perform feature fusion on the channel dimension of integrated energy variables, thereby capturing the nonlinear coupling relationship between different variables.

[0057] In some embodiments, weighted convolution branches are used to capture the short-term dynamics and local changes of the integrated energy system over time.

[0058] In some embodiments, the normalization operation is used to ensure that the gating coefficient is within the range specified in the original text. Within the range.

[0059] In some embodiments, the input data for the model is shown in Table 1.

[0060] Table 1: An example of input data for one model.

[0061] This embodiment dynamically fuses the ChebyshevKAN branch and the weighted convolution branch through a gated residual mechanism (i.e., a gated residual fusion structure), which improves the training stability and generalization ability of the multi-task prediction model while ensuring the feature representation capability.

[0062] In some embodiments, see Figure 4 The multi-task prediction model also includes a long sequence prediction module, which is an Informer network.

[0063] The input structure of the Informer network includes encoder input and decoder input.

[0064] The encoder input is composed of the spatiotemporal fusion features.

[0065] The decoder input is formed by concatenating the spatiotemporal fusion features and placeholder data in chronological order; wherein the placeholder data is filled with zero values.

[0066] The encoder input and decoder input are respectively subjected to embedding representation processing, and the embedding representation is obtained by superimposing numerical feature embedding, position embedding and temporal feature embedding.

[0067] The encoder extracts features from the embedded encoder input through the ProbSparse sparse self-attention mechanism and self-attention distillation operation to obtain encoded features that represent global dependencies.

[0068] The decoder uses the embedded representation as its input to extract temporal-dependent features through a masked multi-head attention mechanism, and interacts with the encoded features through the multi-head attention mechanism to obtain a global temporal feature representation.

[0069] The global temporal features represent the multi-task prediction head used as input to the multi-task prediction model to complete the multivariate prediction task.

[0070] This embodiment effectively reduces the computational complexity of long-term sequence modeling by introducing the sparse attention mechanism of the Informer network, while improving the ability to model long-term dependencies and realizing multi-dimensional feature collaborative learning.

[0071] In some embodiments, the multi-task prediction head is composed of a multi-layer sensing mechanism, with each task prediction head sharing the underlying network parameters and each task having its own unique top-level output layer parameters.

[0072] In some embodiments, the multi-task prediction head is composed of a multilayer perceptron (MLP), and each prediction value (such as a first prediction value, a second prediction value, etc.) is generated by the multi-task prediction head.

[0073] In some embodiments, the multi-task prediction model extracts features from the shared input feature vector using shared parameters to obtain a unified feature representation; the shared input feature vector is composed of the temporal feature representation output by the Informer network; the unified feature representation is used as the input to the multi-task prediction head and is used for feature mapping through different task branches to generate prediction results for each prediction task.

[0074] In some embodiments, the data loss for each forecasting task is measured using mean squared error (MSE). The forecasting tasks refer to forecasting electrical load, thermal load, cooling load, and carbon emissions.

[0075] In some embodiments, in order to adaptively balance the impact of different tasks on the overall optimization objective, an adaptive weighted loss function based on task uncertainty can be introduced, that is, the data loss of each prediction task is dynamically weighted by setting learnable parameters.

[0076] In some embodiments, the method further includes: Obtain the data to be predicted; wherein the data to be predicted includes the current data of electrical load, heat load, cooling load, carbon emissions and weather within the current first preset time period.

[0077] Based on the data to be predicted, a trained multi-task prediction model is invoked to generate future prediction results; wherein, the future prediction results include the fifth predicted value of electrical load, the sixth predicted value of heat load, the seventh predicted value of cooling load, and the eighth predicted value of carbon emissions within the second preset time period in the future.

[0078] In some embodiments, the first preset duration can be 24 hours prior to the current time.

[0079] In some embodiments, the second preset duration can be 24 hours from the current moment.

[0080] This embodiment achieves collaborative prediction of electricity, heat, cooling loads and carbon emissions within a multi-task learning framework, enhancing the information sharing capability among different energy variables. Furthermore, by introducing PINN, carbon emission mechanism constraints are embedded into the model training process, achieving a deep integration of data-driven and physical constraints, thereby improving the accuracy and physical rationality of the prediction results.

[0081] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0082] For ease of understanding, a specific embodiment will be used as an example: Please see Figure 5 The method includes S1 to S6: S1. Collect historical meteorological data, electricity load data, heat load data, cooling load data and carbon emission data of the integrated energy system. Perform missing value imputation and normalization on the data. Perform autocorrelation and cross-correlation analysis on the input variables to select variables with strong correlation and appropriate sliding time windows to form multivariate time series data.

[0083] In practice, carbon emissions, electricity load, cooling load, and heating load data were collected from the integrated energy system of Arizona State University, Tempe Campus, from January to December 2020, and in February and June 2021, with a time resolution of 1 hour. Meteorological data for the corresponding time periods was also acquired, sourced from the U.S. National Solar Radiation Database, including meteorological characteristics such as temperature, precipitation, and air pressure. Furthermore, timestamp information, holiday information, and weekday information were used as input feature variables.

[0084] During data acquisition, some data may be missing or abnormal due to environmental factors, hardware failures, manual input errors, or abnormal data transmission. To address these missing and abnormal values, deletion or linear interpolation methods are used to clean the original data and obtain a complete and reliable data sequence.

[0085] To eliminate the dimensional differences between different feature variables and improve the stability of model training, the cleaned data is normalized. In one embodiment, the Min-Max normalization method is used to map each feature data to the interval [0,1].

[0086] Based on the above correlation analysis, electrical load, heat load, cooling load, carbon emissions, temperature, precipitation, and atmospheric pressure were ultimately selected as the input feature variables for the model.

[0087] In addition, autocorrelation function analysis was performed on the load series to determine the historical window length for time series modeling.

[0088] The sliding time window length was determined by autocorrelation analysis and set to 24 hours. This window length effectively characterizes the daily cyclical variation characteristics and short-term time-series dependencies of the integrated energy system load and carbon emission data, while avoiding the introduction of too much redundant historical information, thereby reducing the computational complexity of the model while ensuring prediction accuracy.

[0089] S2 employs a Kolmogorov–Arnold network (ChebyshevKAN) based on Chebyshev multinomials to fuse features of the variable dimensions, thereby capturing the nonlinear spatial coupling relationship between different variables.

[0090] In practice, the input tensor is first normalized. To ensure the input data satisfies the domain requirement of the Chebyshev polynomial, the hyperbolic tangent function is used to map the input data, normalizing it to the interval [range]. .

[0091] Subsequently, Chebyshev polynomial basis functions are constructed for each normalized input dimension. Each polynomial is obtained by performing an inverse cosine transform on the input variables and then taking the cosine. A learnable coefficient is then assigned to each polynomial. Next, these polynomials of different orders are weighted and summed according to their corresponding coefficients to obtain the contribution of that input dimension to the current output dimension. Finally, the contributions of all input dimensions are summed to form the final result for that output dimension.

[0092] S3. Construct a weighted convolution (Wconv) with a learnable exponential decay weighting mechanism to extract local temporal dynamic features from the multivariate time series.

[0093] In practical implementation, to enable the convolution operator to be applied more effectively to one-dimensional time series signals, the traditional convolution structure is improved by introducing a set of adaptive weight coefficients that vary with time distance before the convolution kernel operation. Since time series data typically exhibit significant local correlations, data from adjacent time steps usually contribute more significantly to the features of the current moment, while historical signals from farther away have a relatively weaker impact on the current prediction. Therefore, by setting weight coefficients that decay exponentially with time distance, the modeling ability of the convolution operator for the temporal features of the time series can be improved.

[0094] Taking electrical load as an example, when the kernel length is... In this case, an adaptive weighting coefficient is introduced. The calculation formula is shown above.

[0095] Furthermore, in order to ensure the stability of the weights at each time step and to avoid the weight values ​​being too large or too small and causing a shift in the convolution result, the weight coefficients are normalized so that the sum of all weight coefficients is 1.

[0096] Furthermore, the convolution kernel parameters and corresponding weight coefficient matrices of the electrical load characteristic channels are constructed.

[0097] Furthermore, by performing element-wise multiplication of the weight coefficients with the kernel parameters, the weighted convolution kernel can be obtained, the expression of which is given above.

[0098] At time step When performing convolution calculations on the electrical load sequence, the expression for its convolution output is given above.

[0099] It should be noted that when the weighting coefficient When all values ​​are 1, the weighted convolution operation described above degenerates into a traditional one-dimensional convolution operation. By introducing time decay weights, the model's focus on recent time information can be enhanced, thereby improving its ability to extract time features from electrical load time series and further improving the expressive power of the prediction model.

[0100] S4. Construct the WCKAN feature extraction module, and fuse the spatial nonlinear features output by the ChebyshevKAN branch with the temporal local dynamic features output by the improved Wconv branch through a gated residual fusion structure to obtain enhanced spatiotemporal fusion features.

[0101] In practice, the ChebyshevKAN branch mainly utilizes the high-order approximation capability of Chebyshev polynomials to perform feature fusion on the channel dimension of comprehensive energy variables, thereby capturing the nonlinear coupling relationship between different variables.

[0102] The Wconv branch is used to capture the short-term dynamics and local changes of integrated energy systems over time.

[0103] To achieve adaptive fusion of the two feature representations, a gating mechanism is introduced into the model. This gating mechanism dynamically adjusts the outputs of the two branches through a trainable gating parameter. Specifically, let the trainable parameter be... The gating coefficients are obtained by mapping them using the Sigmoid activation function.

[0104] The final output of the WCKAN model is achieved by fusing and adding the original input with the outputs of the ChebyshevKAN branch and the improved WConv branch under a dynamic gating mechanism, thereby enhancing and adaptively integrating the original features.

[0105] S5, the enhanced spatiotemporal fusion features are input into the Informer long sequence prediction module, and the long-term sequence dependencies are extracted using the sparse attention mechanism to obtain the global temporal feature representation.

[0106] In practical implementation, during the input phase, Informer maps multi-energy time series data into a unified embedded representation, whose input embedding consists of numerical feature embedding, location embedding and time feature embedding superimposed.

[0107] To improve the efficiency of long sequence modeling, the Informer model introduces a sparse attention mechanism in the attention calculation process. By selectively calculating key queries, it reduces the time complexity of traditional self-attention mechanisms in long sequence calculations, thereby improving computational efficiency while ensuring the model's expressive power.

[0108] To further reduce the computational and storage overhead in long sequence modeling, Informer introduces a self-attention distillation mechanism.

[0109] The sparse attention and self-attention distillation mechanisms described above can effectively capture long-term dependencies in multivariate time series of integrated energy systems and obtain corresponding global time series feature representations, providing high-quality feature inputs for subsequent load forecasting and carbon emission forecasting.

[0110] S6 constructs a multi-task prediction head with shared parameters. During model training, a physical information neural network is introduced to construct carbon emission mechanism constraints as a physical loss function and jointly optimize it with the data-driven loss function. This achieves the integration of physical constraints and data-driven learning, thereby outputting the collaborative prediction results of electricity, heat, cooling load and carbon emissions of the integrated energy system.

[0111] In practice, the shared input features are composed of the temporal feature representations output by the Informer network; a unified feature representation is extracted through shared parameters and used as the input to the multi-task prediction head composed of a multilayer perceptron to generate the prediction results for each task.

[0112] Furthermore, the data loss for each task was measured using mean squared error.

[0113] Furthermore, to adaptively balance the impact of different tasks on the overall optimization objective, an adaptive weighted loss function based on task uncertainty is introduced. Learnable parameters are used to dynamically weight the losses for each task.

[0114] Furthermore, in one embodiment, physical information neural network constraints are introduced to embed the carbon emission mechanism in the integrated energy system into the model training process. In the integrated energy system, the total carbon emissions of the system... It can be considered as being caused by electrical load Cooling load and heating load The result is a result of joint contributions. Based on the assumption of linear superposition of energy conservation and emission factors, an approximate linear coupling relationship can be established, and the calculation formula is shown above.

[0115] Furthermore, physical residuals are constructed during model training, and the calculation formula is shown above.

[0116] Furthermore, the physical loss function is calculated using the weighted mean square error form, the expression of which is given above.

[0117] Finally, by jointly optimizing the data-driven loss and the physical constraint loss, the expression for the model's total loss function is obtained, as shown above.

[0118] By incorporating the carbon emission mechanism of the integrated energy system into the model training process, the model is simultaneously subject to data-driven learning and physical constraints, thereby improving the accuracy of prediction results and enabling coordinated prediction of electrical load, cooling load, heating load, and carbon emissions of the integrated energy system.

[0119] In summary, the integrated energy system multi-energy load and carbon emission co-prediction method that incorporates physical constraints in the above embodiments of the present invention has the following beneficial effects: (1) Construct a Kolmogorov–Arnold network based on Chebyshev polynomials to realize global nonlinear mapping and fusion of multivariate features, effectively characterize the high-order nonlinear spatial coupling relationship between electricity, heat, cooling load and carbon emissions, and improve the multi-energy variable modeling capability.

[0120] (2) The introduction of weighted convolution with learnable exponential decay mechanism enhances the model’s ability to capture local dynamic changes in the time dimension and can more accurately reflect the short-term fluctuation characteristics of load.

[0121] (3) The WCKAN-Informer feature extraction module is proposed. The gated residual mechanism is used to dynamically fuse the ChebyshevKAN branch and the improved WConv branch. While ensuring the feature expression ability, the model training stability and generalization ability are improved. Combined with the sparse attention mechanism of the Informer long sequence prediction model, the computational complexity of long sequence modeling is effectively reduced, while the ability to model long-term dependencies is improved, realizing multi-dimensional feature collaborative learning.

[0122] (4) Under the multi-task learning framework, the collaborative prediction of electricity, heat, cooling load and carbon emissions was realized, which enhanced the information sharing ability among different energy variables. By introducing the physical information neural network, the carbon emission mechanism constraint was embedded into the model training process, realizing the deep integration of data-driven and physical constraints, and improving the accuracy and physical rationality of the prediction results.

[0123] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an information processing method for an integrated energy system.

[0124] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0125] In one embodiment, a computer-readable storage medium is provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps in the above-described method embodiments.

[0126] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0127] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0128] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0129] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An information processing method for integrated energy systems, characterized in that, The information processing method for integrated energy systems includes: In response to the input of the training sample set, the training sample set is input into the multi-task prediction model to obtain a first predicted value of electrical load, a second predicted value of thermal load, a third predicted value of cooling load, and a fourth predicted value of carbon emissions; wherein, the training sample set is constructed based on historical data of electrical load, thermal load, cooling load, carbon emissions, and meteorology. A carbon balance physical constraint loss term is constructed; wherein, the carbon balance physical constraint loss term is used to measure the deviation between the fourth predicted value and the estimated value of carbon emissions, and the estimated value is calculated by the first predicted value, the second predicted value and the third predicted value through a linear coupling relationship; The total loss function is constructed by weighted summing of the carbon balance physical constraint loss term and the data-driven loss term, and then updated by backpropagation of the parameters of the multi-task prediction model.

2. The information processing method for an integrated energy system according to claim 1, characterized in that, The multi-task prediction model includes a weighted convolution and a Chebyshev-Kolmogorov-Arnold (WCKAN) network module. The WCKAN module includes a Chebyshev KAN branch based on Chebyshev multinomials, used for feature fusion of the input tensor to obtain spatial coupling features. These spatial coupling features describe the nonlinear spatial coupling relationships between different feature variables in the input tensor. The process of fusing these spatial coupling features is as follows: The input tensor to the ChebyshevKAN branch is normalized to the interval [-1, 1] using the hyperbolic tangent function. For each feature dimension of the normalized input tensor, a Chebyshev polynomial basis function of the highest order m is constructed, where m is a configurable hyperparameter. Learnable coefficients are assigned to the function values ​​of each order of Chebyshev polynomial, and the function values ​​of each order are weighted and summed to obtain the feature representation corresponding to that dimension. The feature representations of all dimensions of the normalized input tensor are concatenated into vectors to obtain the spatial coupling features.

3. The information processing method for an integrated energy system according to claim 2, characterized in that, The WCKAN module further includes a weighted convolution WConv branch, used to extract the temporal features of each variable in the input tensor; wherein, the input tensor is obtained by preprocessing the training sample set; the process of extracting the temporal features is as follows: The weighted convolution branch sets a set of learnable first weight coefficients for each feature channel. The first weight coefficients are constructed based on the distance between each position of the convolution kernel and the current time step, and decay exponentially as the distance increases. The first weight coefficients are normalized. The normalized first weight coefficients are multiplied element-wise with the standard convolution kernel parameters to obtain a learnable weighted convolution kernel. The weighted convolution kernel is slid along the time dimension with a stride of 1 to perform convolution operations to obtain the time features.

4. The information processing method for an integrated energy system according to claim 3, characterized in that, The WCKAN module further includes a gated residual fusion structure, used to adaptively fuse the spatial coupling features and the temporal features to obtain spatiotemporal fusion features; the process of fusing to obtain the spatiotemporal fusion features is as follows: Initialize the learnable gating parameters; The learnable gating parameters are normalized using the Sigmoid activation function to obtain the gating coefficients. The spatial coupling feature and the temporal feature are added element by element, and then multiplied by the gating coefficient to obtain the weighted fusion feature; The weighted fusion feature is added element-wise to the original feature in the tensor to obtain the spatiotemporal fusion feature.

5. The information processing method for an integrated energy system according to claim 4, characterized in that, The multi-task prediction model also includes a long sequence prediction module, which is an Informer network. The input structure of the Informer network includes encoder input and decoder input; The encoder input is composed of the spatiotemporal fusion features; The decoder input is formed by concatenating the spatiotemporal fusion features and placeholder data in chronological order; wherein the placeholder data is filled with zero values. The encoder input and decoder input are respectively subjected to embedding representation processing, and the embedding representation is obtained by superimposing numerical feature embedding, positional embedding and temporal feature embedding; The encoder extracts features from the embedded encoder input through the ProbSparse sparse self-attention mechanism and self-attention distillation operation to obtain encoded features that characterize global dependencies. The decoder uses the embedded representation as its input to extract temporal-dependent features through a masked multi-head attention mechanism, and interacts with the encoded features through the multi-head attention mechanism to obtain a global temporal feature representation. The global temporal features represent the multi-task prediction head used as input to the multi-task prediction model to complete the multivariate prediction task.

6. The information processing method for an integrated energy system according to claim 1, characterized in that, The multi-task prediction model includes a multi-task prediction head, which is composed of a multi-layer perceptron. Each task prediction head shares the underlying network parameters, while each task has its own unique top-level output layer parameters.

7. The information processing method for an integrated energy system according to claim 1, characterized in that, The linear coupling relationship is specifically as follows: The calculated value is equal to the sum of the learnable error compensation term and the first product, the second product, and the third product; Wherein, the first product is the product of the first predicted value and the learnable electrical carbon emission coefficient, the second product is the product of the second predicted value and the learnable thermal carbon emission coefficient, and the third product is the product of the third predicted value and the learnable cold carbon emission coefficient; the learnable error compensation term and the learnable carbon emission coefficient are jointly optimized with the parameters of the multi-task prediction model through backpropagation during the training process.

8. The information processing method for an integrated energy system according to claim 1, characterized in that, The method further includes: Acquire the data to be predicted; wherein, the data to be predicted includes the current data of electrical load, heat load, cooling load, carbon emissions and weather within the current first preset time period; Based on the data to be predicted, a trained multi-task prediction model is invoked to generate future prediction results; wherein, the future prediction results include the fifth predicted value of electrical load, the sixth predicted value of heat load, the seventh predicted value of cooling load, and the eighth predicted value of carbon emissions within the second preset time period in the future.

9. A computer device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the information processing method for an integrated energy system as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the information processing method for an integrated energy system as described in any one of claims 1-8.

Images