A long-term multi-load forecasting method and system based on parallelized closed-form continuous neural networks

Abstract

The application discloses a long-term multi-load prediction method and system based on a parallel closed continuous neural network, and belongs to the technical field of comprehensive energy system scheduling. The method comprises the following steps: performing autocorrelation analysis on detection data to obtain multiple down-sampling periods; performing down-sampling on the detection data based on the down-sampling periods to obtain sub-sequences; fusing multiple sub-sequences corresponding to the multiple down-sampling periods after weighted summation to obtain fusion features; dividing the fusion features based on time points to obtain multiple time point branch features; inputting the multiple time point branch features into an improved liquid neural network in parallel to obtain multiple updated hidden states; splicing the multiple updated hidden states to obtain spliced time sequence features; and performing high-frequency filtering and decoding on the spliced time sequence features to obtain a prediction result of an output parameter. The application can improve the intelligent level and operation efficiency of comprehensive energy system scheduling, and realizes optimal configuration of accuracy and training cost in a multi-prediction time domain task.

Description

Technical Field

[0001] This application belongs to the field of integrated energy system dispatching technology, specifically relating to a long-term multi-load forecasting method and system based on a parallelized closed continuous neural network. Background Technology

[0002] Traditional load forecasting methods are mainly based on statistical methods such as time series analysis and regression models, or utilize early machine learning algorithms. These methods typically rely on manual feature engineering and have limited ability to model complex nonlinear time series relationships and coupled factors from multiple sources (such as weather and holidays). When dealing with long-term forecasting tasks, they suffer from problems such as insufficient model generalization ability and rapid accumulation of forecasting errors over time.

[0003] In recent years, deep learning-based prediction methods, such as recurrent neural networks and their variants, as well as the Transformer architecture, have been widely used due to their powerful sequence modeling capabilities. However, existing methods still face significant challenges when applied to long-term multi-load prediction. First, models represented by RNNs are limited by sequential computation and the vanishing gradient problem, making it difficult to effectively model long-range temporal dependencies, and their training and inference efficiency is low. Second, although Transformer-type models can capture global dependencies, the computational complexity of their self-attention mechanism increases quadratically with the sequence length, resulting in huge computational overhead when processing long historical windows or multivariate sequences, making it difficult to meet the real-time requirements of scheduling systems. In addition, existing research mostly focuses on short-term prediction, lacking effective lightweight modeling methods for problems such as multi-timescale coupled dynamics, decay of historical information relevance, and amplification of cumulative errors in long-term prediction, making it difficult to achieve a good balance between prediction accuracy and computational efficiency. Summary of the Invention

[0004] Purpose of the invention: This application develops a long-term multi-load prediction method and system based on parallelized closed continuous neural networks, aiming to solve the technical problems in the prior art.

[0005] Technical Solution: In a first aspect, embodiments of this application provide a long-term multi-load forecasting method based on a parallelized closed continuous neural network, comprising: The prediction model is obtained based on a multi-period parallel closed neural network. The input parameters of the prediction model include at least one of three features: energy auxiliary features, time series index features, and meteorological features. The output parameters include load features. The prediction model processes the detection data of the input parameters to obtain the prediction results of the output parameters, including: Autocorrelation analysis was performed on the detection data to obtain multiple downsampling periods; The detection data is downsampled based on the downsampling period to obtain a subsequence; The multiple subsequences corresponding to the multiple downsampling periods are weighted and summed and then fused to obtain the fusion feature of the time series features of multiple periods; Based on the time-division of the fusion features, multiple time-branch features are obtained; The features of multiple time-branch combinations are input in parallel into the improved liquid neural network to obtain multiple updated hidden states; Concatenate multiple updated hidden states to obtain concatenated temporal features; The splicing timing features are subjected to high-frequency filtering and decoding to obtain the prediction results of the output parameters.

[0006] In some embodiments, the step of fusing multiple subsequences corresponding to multiple downsampling periods after weighted summation to obtain fused features of multiple periodic time-series features includes: Determine the prediction time length and combine it with a shared parameter model to predict the subsequence, thereby obtaining the predicted sequence; Upsample the predicted sequence back to an upsampled sequence with a length equal to the predicted time length; The time-series features are obtained by weighted summation of multiple upsampled sequences corresponding to the downsampling period. By fusing the temporal features corresponding to multiple downsampling periods, a fused feature containing temporal features of multiple periods is obtained.

[0007] In some embodiments, before the step of downsampling the detection data based on the downsampling period to obtain a subsequence, the method further includes: The detection data is aggregated based on one-dimensional convolutional sliding aggregation, and adjacent data points in the detection data are weighted and fused; the characterization formula of the sliding aggregation is as follows: ; in, To analyze the test data The sequence after sliding aggregation; It is a one-dimensional convolution.

[0008] In some embodiments, the step of obtaining the improved liquid neural network includes: Obtaining closed-form approximate solutions for liquid neural networks; The hidden states of the improved liquid neural network are obtained by replacing the exponential decay term in the closed-form approximation solution with the inverse sigmoid function; the hidden state representation formula of the improved liquid neural network is as follows: ; in, and For liquid neural networks Learnable functions; It is the inverse sigmoid function; The hidden states of the improved liquid neural network; For a specific moment; Branching features at time points; For liquid neural networks The set of all trainable weights and biases within the ; It is the Hadamard product.

[0009] In some embodiments, before the step of splicing multiple updated hidden states to obtain splicing temporal features, the method further includes: Determine the processing cycle that includes multiple time points; The time branch features at the same time within multiple consecutive processing cycles with lag are input into the improved liquid neural network to obtain the hidden states updated after multiple lag times, and then weighted and summed proportionally.

[0010] In some embodiments, the step of performing high-frequency filtering and decoding on the spliced ​​time-series features to obtain load prediction results includes: The splicing time sequence features are normalized to obtain a normalized sequence; Perform a real-signal Fourier transform on the normalized sequence along the time dimension to obtain the first transform data; The first transformed data is low-pass filtered to retain low-frequency components and cut off high-frequency noise to obtain filtered data. The filtered data is upsampled based on a single linear layer, the dimension of the filtered data is expanded and interpolated to obtain the upsampled data. The upsampled data is filled into the corresponding target frequency domain dimension based on a predetermined prediction time length to obtain filled frequency domain data; Perform an inverse Fourier transform on the filled frequency domain data and adjust the amplitude to obtain the second transform data; The second transformed data is inversely normalized, and the predicted portion is extracted as the load forecast value.

[0011] In some embodiments, the step of obtaining the input parameters includes: Determine the undetermined input parameters and output parameters, and obtain the operating data of the undetermined input parameters and output parameters. The undetermined input parameters include energy auxiliary characteristics, time series index characteristics and meteorological characteristics. The load characteristics of the output parameters include electrical load, cooling load and / or heating load. Based on the operational data, obtain the Pearson correlation coefficient between the undetermined input parameter and the output parameter; The input parameters are determined from the undetermined input parameters based on the Pearson correlation coefficient.

[0012] In some embodiments, the step of obtaining the runtime data includes: Obtain the original data of the undetermined input and output parameters; The anomaly range of the original data is determined by the quartile method, and the original data within the anomaly range is removed to obtain the first processed data. The first processed data is processed based on the global three standard deviation principle to obtain the second processed data; The second processed data is processed based on the principle of three standard deviations locally to obtain the running data.

[0013] In some embodiments, the step of determining the outlier range of the original data using the quartile method includes: Determine the first and third quartiles of the original data; The interquartile distance is determined based on the first quartile and the third quartile, and the formula for representing the interquartile distance is as follows: ; in, Interquartile distance; It is the third quartile; It is the first quartile; The lower and upper limits of normal values ​​are obtained based on the interquartile range; the formulas for representing the lower and upper limits of normal values ​​are as follows: ; ; in, This is the lower limit threshold of the normal value; This is the upper limit threshold of the normal value; The range of abnormal values ​​is obtained based on the lower limit threshold and the upper limit threshold of normal values; The process of processing the first processed data based on the global three-standard-deviation principle to obtain the second processed data includes: A first normal range is obtained based on the mean and standard deviation of the first processed data; the formula for representing the first normal range is as follows: ; ; ; in, This is within the first normal range; For the first The first processing data of each parameter The mean, For a moment, The total number of moments; For the first The first processing data of each parameter Standard deviation; The outlier handling function removes positive values ​​outside the first normal range to obtain the second processed data; The process of processing the second processed data based on the principle of three local standard deviations to obtain the running data includes: Define the sliding window; Local data is obtained by sliding the sliding window over the second processed data. A second normal range is obtained based on the mean and standard deviation of the local data. The formula for representing the second normal range is as follows: ; ; ; ; ; in, This is the upper limit threshold of the second normal range; This is the lower limit threshold of the second normal range; For sliding windows Inner The second processing data of each parameter The mean; For sliding windows Inner The second processing data of each parameter Standard deviation; For sliding windows Internal second processing data A set of time indexes; For time indexing; For adjustment factors; Remove data from the second processed data that falls outside the second normal range, and obtain the running data.

[0014] Secondly, embodiments of this application also provide a long-term multi-load forecasting system based on a parallelized closed continuous neural network, comprising: The model acquisition module is used to acquire a prediction model based on a multi-period parallel closed neural network. The input parameters of the prediction model include at least one of three features: energy auxiliary features, time series index features, and meteorological features. The output parameters include load features. The load forecasting module is used to process the detection data of the input parameters through the forecasting model to obtain the forecasting results of the output parameters; the forecasting model includes: A multi-period sparse feature fusion submodule is used to perform autocorrelation analysis on the detection data to obtain multiple downsampling periods; downsample the detection data based on the downsampling periods to obtain subsequences; and fuse the multiple subsequences corresponding to the multiple downsampling periods after weighted summation to obtain fusion features of multiple periodic time-series features. A parallel closed continuous neural network submodule is used to divide the fusion features based on time points to obtain multiple time-based branch features; input the multiple time-based branch features in parallel into the improved liquid neural network to obtain multiple updated hidden states; and concatenate the multiple updated hidden states to obtain concatenated temporal features. The frequency domain filtering submodule is used to perform high-frequency filtering and decoding on the spliced ​​timing features to obtain the prediction results of the output parameters.

[0015] Beneficial Effects: Compared with the prior art, the present application provides a long-term multi-load forecasting method based on a parallel closed continuous neural network, including obtaining a forecasting model based on a multi-period parallel closed neural network. The input parameters of the forecasting model include at least one of three features: energy auxiliary features, time series index features, and meteorological features. The output parameters include load features. The forecasting model processes the detection data of the input parameters to obtain the forecasting results of the output parameters, including: performing autocorrelation analysis on the detection data to obtain multiple downsampling periods; downsampling the detection data based on the downsampling periods to obtain subsequences; and fusing the multiple subsequences corresponding to the multiple downsampling periods after weighted summation to obtain the fusion feature of the time series features of multiple periods, simplifying the original sequence forecasting task into a cross-period trend forecasting task. This method enables the model to predict load data over longer periods, achieving efficient decomposition and extraction of cyclical features from multiple load data periods under constraints of parameter quantity and computational cost. Based on time-division feature fusion, multiple time-branch features are obtained. These multiple time-branch features are input in parallel into an improved liquid neural network, approximating the solution of differential equations through closed-form expressions. Parallelism is introduced to improve computational efficiency while maintaining accuracy, thus balancing accuracy and computational cost across different prediction durations, and obtaining multiple updated hidden states. These updated hidden states are then concatenated to obtain concatenated time-series features. High-frequency filtering and decoding of the concatenated time-series features, combined with frequency domain filtering to remove high-frequency noise, overcomes the limitations of existing methods that are mostly limited to short-term predictions, obtaining prediction results for output parameters. This application can improve the intelligence level and operational efficiency of integrated energy system scheduling from the data source, achieving an overall optimal configuration of accuracy and training cost in multi-prediction time-domain tasks. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the steps of the long-term multi-load forecasting method based on a parallelized closed continuous neural network provided in this application embodiment; Figure 2 A flowchart illustrating the steps for obtaining prediction results of output parameters in the long-term multi-load prediction method based on parallelized closed continuous neural networks provided in this application embodiment; Figure 3 A flowchart illustrating the steps of obtaining fused features of multiple periodic time series features in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment; Figure 4 A flowchart illustrating the steps of high-frequency filtering and decoding of spliced ​​time-series features in the long-term multi-load prediction method based on parallelized closed continuous neural networks provided in this application embodiment; Figure 5 A flowchart illustrating the steps for obtaining input parameters in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment; Figure 6 A flowchart illustrating the steps for obtaining operational data in the long-term multi-load forecasting method based on a parallelized closed continuous neural network provided in this application embodiment; Figure 7 A flowchart illustrating the steps of determining the anomaly range of the original data using the quartile method in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment; Figure 8 A flowchart illustrating the steps of processing the second processing data based on the principle of local three standard deviations in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment; Figure 9 A module connection diagram of a long-term multi-load forecasting system based on a parallelized closed continuous neural network provided in an embodiment of this application; Figure 10 The flowchart illustrates the process of constructing a prediction model from raw data containing load characteristics, meteorological characteristics, and time index characteristics in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment. Figure labels: 10, Model acquisition module; 20, Load prediction module; 21, Multi-period sparse feature fusion sub-module; 22, Parallel closed continuous neural network sub-module; 23, Frequency domain filtering sub-module. Detailed Implementation

[0018] 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 a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0019] Traditional load forecasting methods are mainly based on statistical methods such as time series analysis and regression models, or utilize early machine learning algorithms. These methods typically rely on manual feature engineering and have limited ability to model complex nonlinear time series relationships and coupled factors from multiple sources (such as weather and holidays). When dealing with long-term forecasting tasks, they suffer from problems such as insufficient model generalization ability and rapid accumulation of forecasting errors over time.

[0020] In recent years, deep learning-based prediction methods, such as recurrent neural networks and their variants, as well as the Transformer architecture, have been widely used due to their powerful sequence modeling capabilities. However, existing methods still face significant challenges when applied to long-term multi-load prediction. First, models represented by RNNs are limited by sequential computation and the vanishing gradient problem, making it difficult to effectively model long-range temporal dependencies, and their training and inference efficiency is low. Second, although Transformer-type models can capture global dependencies, the computational complexity of their self-attention mechanism increases quadratically with the sequence length, resulting in huge computational overhead when processing long historical windows or multivariate sequences, making it difficult to meet the real-time requirements of scheduling systems. In addition, existing research mostly focuses on short-term prediction, lacking effective lightweight modeling methods for problems such as multi-timescale coupled dynamics, decay of historical information relevance, and amplification of cumulative errors in long-term prediction, making it difficult to achieve a good balance between prediction accuracy and computational efficiency.

[0021] In view of this, embodiments of this application provide a long-term multi-load forecasting method based on a parallelized closed-loop continuous neural network, including obtaining a forecasting model based on a multi-period parallel closed-loop neural network. The input parameters of the forecasting model include at least one of three features: energy auxiliary features, time-series index features, and meteorological features. The output parameters include load features, which include electrical load, cooling load, and / or heating load. The forecasting model processes the detection data of the input parameters to obtain the forecasting results of the output parameters, including: performing autocorrelation analysis on the detection data to obtain multiple downsampling periods; downsampling the detection data based on the downsampling periods to obtain subsequences; fusing the multiple subsequences corresponding to the multiple downsampling periods after weighted summation to obtain fused features of multiple periodic time-series features; dividing the fused features based on time to obtain multiple time-branch features; inputting the multiple time-branch features in parallel into an improved liquid neural network to obtain multiple updated hidden states; concatenating the multiple updated hidden states to obtain concatenated time-series features; and performing high-frequency filtering and decoding on the concatenated time-series features to obtain the forecasting results of the output parameters.

[0022] This application aims at parameter simplification. It designs a multi-period sparse feature fusion submodule in the prediction model based on a channel-independent strategy, achieving efficient decomposition and extraction of periodic features from multi-load data under limited parameter quantity and computational overhead. A frequency domain filtering submodule is introduced, operating only on low-frequency components and using a single-layer linear layer for frequency domain interpolation, achieving lightweight prediction with extremely low parameter quantity and effectively solving the problem of high computational complexity in existing composite models. The optimal period combination is selected through autocorrelation analysis to construct a multi-period sparse feature fusion module, simplifying the original sequence prediction task into cross-period trend prediction, enabling the model to predict load data over longer periods. Frequency domain filtering is combined to remove high-frequency noise, overcoming the limitation of existing methods being mostly limited to short-term prediction. A parallel closed-loop continuous neural network submodule is proposed, fusing recurrent neural networks and continuous-time neural networks to construct a parallel closed-loop continuous neural network module. This module approximates the solution of differential equations through closed-loop expressions, introducing a parallel mechanism to improve computational efficiency while maintaining accuracy, thus balancing accuracy and computational overhead across different prediction durations. This application can improve the intelligence level and operational efficiency of integrated energy system scheduling from the data source, and achieves the overall optimal configuration of accuracy and training cost in multi-prediction time domain tasks.

[0023] In some embodiments, please refer to Figure 1 and Figure 10 , Figure 1 This is a flowchart illustrating the steps of the long-term multi-load forecasting method based on a parallelized closed continuous neural network provided in this application embodiment. Figure 10The flowchart illustrates the process of constructing a prediction model from raw data containing load characteristics, meteorological characteristics, and time index characteristics in the long-term multiple load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment. Specifically, the long-term multiple load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment is implemented through steps 100 to 200: Step 100: Obtain the prediction model based on a multi-period parallel closed neural network.

[0024] In some embodiments, the input parameters of the prediction model include at least one of three features: energy-assisted features, time-series index features, and meteorological features. The output parameters include electrical load, cooling load, and / or heating load. Specifically, meteorological features include temperature, humidity, wind speed, and light intensity. Energy-assisted features include a comprehensive energy index and greenhouse gas emissions. The comprehensive energy index is defined as the cumulative value of various energy sources after being uniformly converted into heat units. The formula for calculating the comprehensive energy index is as follows: ; in, For electrical load, For cooling load, For heat load, electrical load is measured in kilowatts (kW), cooling load is measured in tons of cooling per hour (Ton / h), and heat load is expressed in millions of British thermal units per hour (MMBTU / h). The conversion relationships between the load units are as follows: .

[0025] Understandably, greenhouse gas emissions (GHG) quantify the level of greenhouse gas emissions caused by integrated energy use. The introduction of these two energy-ancillating features injects integrated energy information deeply relevant to the prediction task into the model with zero additional acquisition cost, effectively improving the model's response and expressive performance to changes in multi-energy loads.

[0026] In some embodiments, please refer to Figure 5 , Figure 5 This document provides a flowchart of the steps for obtaining input parameters in a long-term multi-load forecasting method based on a parallelized closed continuous neural network, as provided in this application embodiment. The method for obtaining input parameters in this application embodiment is specifically implemented through steps 110 to 130: Step 110: Determine the input and output parameters to be determined, and obtain the running data of the input and output parameters to be determined.

[0027] In some embodiments, please refer to Figure 6 , Figure 6This is a flowchart illustrating the steps for obtaining operational data in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment. The method for obtaining operational data in this application embodiment is specifically implemented through steps 111 to 114: Step 111: Obtain the raw data of the input and output parameters to be determined.

[0028] In some embodiments, after acquiring the raw data, a data outlier handling function is defined. The value of each parameter at each time step Values ​​greater than 300,000 or negative values ​​are set to null; outlier handling function. The characterization formula is as follows: ; in, The original data for the i-th parameter at time t. A feature index, used to characterize a specific parameter. This is the time index; 300000 is the set upper limit threshold for normal values. Positive values ​​exceeding this value are considered abnormal and removed. Step 112: Determine the outlier range of the original data using the quartile method, and remove the original data within the outlier range to obtain the first processed data.

[0029] In some embodiments, please refer to Figure 7 , Figure 7 The flowchart illustrates the steps of determining the anomaly range of the original data using the quartile method in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment. Specifically, the method for determining the anomaly range of the original data using the quartile method in this application embodiment is implemented through steps 1121 to 1124: Step 1121: Determine the first and third quartiles of the original data.

[0030] Step 1122: Determine the interquartile distance based on the first quartile and the third quartile.

[0031] In some embodiments, the formula for representing the interquartile range includes: ; in, Interquartile distance; It is the third quartile, in this application It is the 85th percentile. ; As the first quartile, in this application It is the 15th percentile. .

[0032] Step 1123: Obtain the lower limit threshold and upper limit threshold of normal values ​​based on the interquartile range.

[0033] In some embodiments, the formulas for representing the lower limit threshold and the upper limit threshold of normal values ​​include: ; ; in, This is the lower limit threshold of the normal value; This represents the upper limit threshold of the normal value.

[0034] Step 1124: Obtain the range of outliers based on the lower limit threshold and the upper limit threshold of normal values.

[0035] Step 113: Process the first data based on the global three standard deviation principle to obtain the second data.

[0036] In some embodiments, this application first obtains a first normal range based on the mean and standard deviation of the first processed data, and the formula for representing the first normal range includes: ; ; ; in, This is within the first normal range; For the first The first processing data of each parameter The mean, For a moment, The total number of moments; For the first The first processing data of each parameter Standard deviation; Then, based on the outlier handling function, positive values ​​outside the first normal range are removed to obtain the second processed data.

[0037] Step 114: Process the second processing data based on the principle of three standard deviations locally to obtain the running data.

[0038] In some embodiments, please refer to Figure 8 , Figure 8 This is a flowchart illustrating the steps of processing the second processed data based on the local three-standard-deviation principle in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment. Specifically, the method for processing the second processed data based on the local three-standard-deviation principle in this application embodiment is implemented through steps 1141 to 1144: Step 1141: Determine the sliding window.

[0039] Step 1142: Obtain local data by sliding a sliding window over the second processed data.

[0040] Step 1143: Obtain the second normal range based on the mean and standard deviation of the local data.

[0041] In some embodiments, the characterization formula for the second normal range includes: ; ; ; ; ; in, This is the upper limit threshold of the second normal range; This is the lower limit threshold of the second normal range; For sliding windows Inner The second processing data of each parameter The mean; For sliding windows Inner The second processing data of each parameter Standard deviation; For sliding windows Internal second processing data A set of time indexes; For time indexing; This is an adjustment factor.

[0042] Step 1144: Remove data from the second processing data that is outside the second normal range and obtain the running data.

[0043] In some embodiments, the input parameters to be determined include energy-aided features, time-series index features, and meteorological features.

[0044] Step 120: Obtain the Pearson correlation coefficient between the undetermined input parameters and the output parameters based on the running data.

[0045] Step 130: Determine the input parameters from the undetermined input parameters based on the Pearson correlation coefficient.

[0046] In some embodiments, after outlier processing, the Pearson correlation coefficient is used to measure the linear correlation strength between parameters, while the Spearman rank correlation coefficient is combined to assess the monotonic trend, thus performing a comprehensive evaluation of the sample characteristics. This application calculates the Pearson correlation coefficient between each input parameter (meteorological characteristics, energy auxiliary characteristics, and time-series index characteristics) and each output parameter (heat, electricity, and cooling loads). Using a correlation coefficient greater than 0.3 as a threshold, based on the correlation analysis results, energy auxiliary characteristics and loads, as well as time-series index characteristics and loads, show high correlations and are retained; however, some meteorological parameters have low correlations with loads, therefore, the low-correlation meteorological characteristics are discarded, completing the selection of input features.

[0047] In some embodiments, after determining the input and output parameters, a multi-period parallel closed neural network is constructed. This multi-period parallel closed neural network includes a cascaded multi-period sparse feature fusion submodule, a parallel closed continuous neural network submodule, and a frequency domain filtering submodule. Then, the historical data of the input parameters processed in steps 111 to 114 are used as the input, and the historical data corresponding to the output parameters are used as the output. The multi-period parallel closed neural network is trained to obtain a prediction model. Specifically, during training, the initial learning rate is set to 0.001, the decay coefficient is 0.98, and the learning rate is updated every three training iterations, for a total of 200 training iterations. The Adam algorithm is used for optimization during training, and the network parameters are updated via backpropagation.

[0048] Step 200: Process the detection data of the input parameters using a prediction model.

[0049] In some embodiments, please refer to Figure 2 , Figure 2 This is a flowchart illustrating the steps for obtaining the prediction results of output parameters in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment. Specifically, the method for obtaining the prediction results of output parameters in this application embodiment is implemented through steps 210 to 270: Step 210: Perform autocorrelation analysis on the detection data to obtain multiple downsampling periods.

[0050] In some embodiments, this application determines the optimal combination of downsampling periods through autocorrelation analysis, specifically selecting four downsampling periods: 3 hours, 6 hours, 12 hours, and 24 hours.

[0051] In some embodiments, before downsampling the detection data based on the downsampling period to obtain subsequences, the long-term multi-load prediction method based on parallelized closed continuous neural networks provided in this application provides a method for one-dimensional convolutional sliding aggregation of detection data and weighted fusion of adjacent data points in the detection data. The characterization formula for sliding aggregation includes: ; in, To analyze the test data The sequence after sliding aggregation; It is a one-dimensional convolution.

[0052] Understandably, this application performs sliding aggregation on the data based on one-dimensional convolution. By weighted fusion of adjacent data points, it can suppress outlier interference and enhance the interactive information between different periods.

[0053] Step 220: Downsample the detection data based on the downsampling period to obtain subsequences.

[0054] Understandably, since the data of each parameter have significant temporal patterns, this application designs a downsampling period to downsample the time series to obtain multiple subsequences, so that prediction can be performed on each subsequence separately, thus transforming the original time series prediction problem into modeling the cross-period evolution trend.

[0055] Step 230: After weighted summation of multiple subsequences corresponding to multiple downsampling periods, fuse them to obtain the fused features of multiple periodic time series features.

[0056] In some embodiments, please refer to Figure 3 , Figure 3 The flowchart illustrates the steps for obtaining the fused features of multiple periodic time series features in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment. Specifically, the method for obtaining the fused features of multiple periodic time series features in this application embodiment is implemented through steps 231 to 234: Step 231: Determine the prediction time length and combine it with the shared parameter model to predict the subsequence and obtain the prediction sequence.

[0057] Understandably, this application employs a channel-independent approach, allowing each parameter's subsequence to learn a shared mapping function, in order to specifically simplify the multi-period sparse feature fusion submodule and decompose multivariate prediction into multiple univariate prediction problems.

[0058] Step 232: Upsample the predicted sequence back to an upsampled sequence with a length equal to the predicted time length.

[0059] Step 233: Perform a weighted summation of multiple upsampled sequences corresponding to the downsampling period to obtain the temporal features.

[0060] In some embodiments, it is assumed that the detection data has a downsampling period. Then the detection data will be downsampled to A length of The subsequence; then a shared parameter model is used to predict the length of the subsequence. The subsequence, and re-encode The length of each subsequence upsampling return is The subsequence. Therefore, for a single downsampling period, the predicted sequence output by the sparse feature fusion submodule is as follows: ; in, The duration is The predicted sequence is used to characterize the first load forecast value; This is a concatenation function; To take the average function; Refers to the temporal position at the beginning of each subsequence; The length of the subsequence; The downsampling period; The sequence length of the detection data to be sampled; For subsequences; The neural network corresponding to the sparse feature fusion submodule; For length is The downsampled subsequence. At this point, the number of parameters in the sparse feature fusion submodule is... .

[0061] Step 234: Fuse the temporal features corresponding to multiple downsampling periods to obtain a fused feature containing temporal features from multiple periods.

[0062] Understandably, the first load prediction value corresponding to the downsampling period obtained by the sparse feature fusion submodule in this application... This can be output as a coarse load forecast result. To obtain a more accurate load forecast result, this application further processes the data using a parallel closed continuous neural network submodule and a frequency domain filtering submodule. To obtain more accurate load forecast results.

[0063] Step 240: Based on the time-separated fusion features, obtain multiple time-separated branch features.

[0064] Step 250: Input the branch features from multiple time steps into the improved liquid neural network in parallel to obtain multiple updated hidden states.

[0065] In some embodiments, when obtaining the improved liquid neural network, the closed-form approximate solution of the liquid neural network is first obtained. The liquid neural network is a typical continuous neural network based on ordinary differential equations (ODEs). The liquid neural network and its closed-form approximate solution calculation formula are defined as follows: ; in, for The hidden state of the liquid neural network at any given time; for The external input to the time equation, in this application for Timing branch characteristics at time points; It is a time constant parameter; For liquid neural networks, For liquid neural networks The set of all trainable weights and biases within the ; Initially hidden state, and for 3D system vector.

[0066] For this formula, formally each time step Place has Hidden state of each hidden unit It can be displayed as follows: ; in, for The corresponding parameter vector, assuming external input For each moment Dimensional input, for The hidden state of the liquid neural network at any given time; Then, by replacing the exponential decay term in the closed-form approximation of the liquid neural network with the inverse sigmoid function to prevent gradient vanishing, the hidden state of the improved liquid neural network is obtained. The hidden state representation formula of the improved liquid neural network includes: ; in, and For liquid neural networks Learnable functions; It is the inverse sigmoid function; The hidden states of the improved liquid neural network; For a specific moment; Branching features at time points; For liquid neural networks The set of all trainable weights and biases within the ; It is the Hadamard product.

[0067] Step 260: Concatenate multiple updated hidden states to obtain concatenated temporal features.

[0068] In some embodiments, a typical cycle of 24 times a day is extracted. The input is combined into a feature matrix. Periodicity The feature matrix at time step is , Hide state at all times Update to the hidden state within a typical cycle. At this point, the output of the module after concatenation is: ; in, This is for splicing temporal features.

[0069] In some embodiments, before concatenating multiple updated hidden states, this application first determines a processing cycle containing multiple time points. Then, the time branch features of the same time point within multiple consecutive processing cycles with lags are input into the improved liquid neural network to obtain the hidden states updated at multiple lag times, and then weighted and summed proportionally. Specifically, this application selects 24 hours as the processing cycle and selects the hidden states of the same time point on the following one, two, and three days, and weights and sums them proportionally during the update, enabling the model to recall historical information from the same period and improving its ability to capture periodic patterns.

[0070] In some embodiments, the formula for selecting the lag time variable is as follows: ; exist During the update, if Greater than respectively In hidden features Before each input into the neural network, a weighted sum is performed according to a certain ratio, corresponding to the same time period before the previous number of days. ,set up Using the index of past days, the final output of this module after concatenation is: ; The calculation results of the parallel closed continuous neural network module will be passed to the frequency domain filtering module.

[0071] Step 270: Perform high-frequency filtering and decoding on the spliced ​​time sequence features to obtain the prediction results of the output parameters.

[0072] In some embodiments, please refer to Figure 4 , Figure 4 This is a flowchart illustrating the steps of high-frequency filtering and decoding of spliced ​​time-series features in the long-term multi-load forecasting method based on parallelized closed continuous neural networks provided in this application embodiment. Specifically, the method for high-frequency filtering and decoding of spliced ​​time-series features in this application embodiment is implemented through steps 271 to 277: Step 271: Normalize the spliced ​​temporal features to obtain a normalized sequence.

[0073] Understandably, this application normalizes the splicing temporal features to eliminate the influence of dimensions, resulting in a normalized sequence. The calculation formula is as follows: ; ; ; in, For normalized sequences; For normalized sequence The mean value obtained by averaging all features; for The variance of each feature sequence; This is for splicing temporal features.

[0074] Step 272: Perform a real signal Fourier transform on the normalized sequence along the time dimension to obtain the first transform data.

[0075] In some embodiments, the representation formula for the real signal Fourier transform of the normalized sequence along the time dimension includes: ; in, The first transformed data is the signal data converted to the frequency domain. For Fast Fourier Transform; The computational dimension of the Fast Fourier Transform; express Dimensions in the frequency domain.

[0076] Step 273: Perform low-pass filtering on the first transformed data, retaining low-frequency components and truncating high-frequency noise to obtain filtered data.

[0077] In some embodiments, the first transformed data is low-pass filtered, retaining the first D low-frequency components. The characterization formula for low-pass filtering the first transformed data includes: ; in, For filtered data; This is a low-pass filter function; its value is 1 when the index is less than D, and 0 otherwise.

[0078] Step 274: Upsample the filtered data based on a single linear layer, expand the dimension of the filtered data and interpolate to obtain the upsampled data.

[0079] In some embodiments, the step of upsampling the filtered data based on a single linear layer includes: ; ; ; ; in, Here, represents the upsampled data, and represents the upsampled frequency domain signal. It is an upsampling function; These are learnable weights; For learnable bias; Low-frequency components Upsampling ratio The expanded components.

[0080] Step 275: Fill the upsampled data into the corresponding target frequency domain dimension based on the predetermined prediction time length, and obtain the filled frequency domain data.

[0081] In some embodiments, based on the prediction time length, the upsampled frequency domain signal is filled into the corresponding target frequency domain dimension to obtain filled frequency domain data, preparing for the inverse transform. The established time series length is... time series Requires target frequency domain dimension .Will Fill to length The calculation formula is as follows: ; in, To fill the frequency domain data; This is a fill function used to... Parts that are not long enough are padded with 0.

[0082] Step 276: Perform an inverse Fourier transform on the filled frequency domain data and adjust the amplitude to obtain the second transform data.

[0083] In some embodiments, performing an inverse Fourier transform on the filled frequency domain data and adjusting the amplitude representation formula includes: ; in, This is the second transformed data; This is the inverse Fourier transform.

[0084] Step 277: Perform inverse normalization on the second transformed data and extract the predicted portion as the load forecast value.

[0085] In some embodiments, denormalization The predicted portion is extracted to obtain the load forecast value, whose characterization formula includes: ; in, This is the second load forecast value, compared to the first load forecast value. It has higher prediction accuracy.

[0086] Understandably, the long-term multi-load forecasting method based on a parallelized closed-loop continuous neural network provided in this application includes obtaining a forecasting model based on a multi-period parallel closed-loop neural network. The input parameters of the forecasting model include at least one of three features: energy auxiliary features, time-series index features, and meteorological features. The output parameters include load features, which include electrical load, cooling load, and / or heating load. The forecasting model processes the detection data of the input parameters to obtain the forecasting results of the output parameters, including: performing autocorrelation analysis on the detection data to obtain multiple downsampling periods; downsampling the detection data based on the downsampling periods to obtain subsequences; fusing the multiple subsequences corresponding to the multiple downsampling periods after weighted summation to obtain fusion features of multiple periodic time-series features; dividing the fusion features based on time to obtain multiple time-branch features; inputting the multiple time-branch features in parallel into an improved liquid neural network to obtain multiple updated hidden states; concatenating the multiple updated hidden states to obtain concatenated time-series features; and performing high-frequency filtering and decoding on the concatenated time-series features to obtain the forecasting results of the output parameters.

[0087] This application aims at parameter simplification. It designs a multi-period sparse feature fusion submodule in the prediction model based on a channel-independent strategy, achieving efficient decomposition and extraction of periodic features from multi-load data under limited parameter quantity and computational overhead. A frequency domain filtering submodule is introduced, operating only on low-frequency components and using a single-layer linear layer for frequency domain interpolation, achieving lightweight prediction with extremely low parameter quantity and effectively solving the problem of high computational complexity in existing composite models. The optimal period combination is selected through autocorrelation analysis to construct a multi-period sparse feature fusion module, simplifying the original sequence prediction task into cross-period trend prediction, enabling the model to predict load data over longer periods. Frequency domain filtering is combined to remove high-frequency noise, overcoming the limitation of existing methods being mostly limited to short-term prediction. A parallel closed-loop continuous neural network submodule is proposed, fusing recurrent neural networks and continuous-time neural networks to construct a parallel closed-loop continuous neural network module. This module approximates the solution of differential equations through closed-loop expressions, introducing a parallel mechanism to improve computational efficiency while maintaining accuracy, thus balancing accuracy and computational overhead across different prediction durations. This application can improve the intelligence level and operational efficiency of integrated energy system scheduling from the data source, and achieves the overall optimal configuration of accuracy and training cost in multi-prediction time domain tasks.

[0088] Accordingly, embodiments of this application also provide a long-term multi-load forecasting system based on a parallelized closed continuous neural network. Please refer to [link to relevant documentation]. Figure 9 , Figure 9This is a module connection diagram of the long-term multi-load forecasting system based on a parallelized closed continuous neural network provided in this application embodiment. The long-term multi-load forecasting system based on a parallelized closed continuous neural network provided in this application embodiment includes: Model acquisition module 10 is used to acquire a prediction model based on a multi-period parallel closed neural network. The input parameters of the prediction model include energy auxiliary features, time series index features and meteorological features, and the output parameters include load features. Load forecasting module 20 is used to process the detection data of input parameters through a forecasting model and obtain the forecast results of output parameters. The forecasting model includes: The multi-period sparse feature fusion submodule 21 is used to perform autocorrelation analysis on the detection data to obtain multiple downsampling periods; to downsample the detection data based on the downsampling periods to obtain subsequences; and to fuse the multiple subsequences corresponding to the multiple downsampling periods after weighted summation to obtain the fusion feature of the time series features of multiple periods. Parallel closed continuous neural network submodule 22 is used to divide and fuse features based on time step to obtain multiple time step branch features; input the multiple time step branch features into the improved liquid neural network in parallel to obtain multiple updated hidden states; and concatenate the multiple updated hidden states to obtain concatenated temporal features. Frequency domain filtering submodule 23 is used to perform high-frequency filtering and decoding on the spliced ​​timing features to obtain the prediction results of the output parameters.

[0089] This application has provided a detailed description of a long-term multi-load prediction method and system based on a parallelized closed continuous neural network, as provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and its core ideas. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A long-term multi-load forecasting method based on parallelized closed continuous neural networks, characterized in that, include: The prediction model is obtained based on a multi-period parallel closed neural network. The input parameters of the prediction model include at least one of three features: energy auxiliary features, time series index features, and meteorological features. The output parameters include load features. The prediction model processes the detection data of the input parameters to obtain the prediction results of the output parameters, including: Autocorrelation analysis was performed on the detection data to obtain multiple downsampling periods; The detection data is downsampled based on the downsampling period to obtain a subsequence; The multiple subsequences corresponding to the multiple downsampling periods are weighted and summed and then fused to obtain the fusion feature of the time series features of multiple periods; Based on the time-division of the fusion features, multiple time-branch features are obtained; The features of multiple time-branch combinations are input in parallel into the improved liquid neural network to obtain multiple updated hidden states; Concatenate multiple updated hidden states to obtain concatenated temporal features; The splicing timing features are subjected to high-frequency filtering and decoding to obtain the prediction results of the output parameters.

2. The long-term multi-load forecasting method based on parallelized closed continuous neural networks according to claim 1, characterized in that, The step of fusing multiple subsequences corresponding to multiple downsampling periods after weighted summation to obtain fused features of multiple periodic time series features includes: Determine the prediction time length and combine it with a shared parameter model to predict the subsequence, thereby obtaining the predicted sequence; Upsample the predicted sequence back to an upsampled sequence with a length equal to the predicted time length; The time-series features are obtained by weighted summation of multiple upsampled sequences corresponding to the downsampling period. By fusing the temporal features corresponding to multiple downsampling periods, a fused feature containing temporal features of multiple periods is obtained.

3. The long-term multi-load forecasting method based on parallelized closed continuous neural networks according to claim 1, characterized in that, Before the step of downsampling the detection data based on the downsampling period to obtain a subsequence, the method further includes: The detection data is aggregated based on one-dimensional convolutional sliding aggregation, and adjacent data points in the detection data are weighted and fused; the characterization formula of the sliding aggregation is as follows: ； in, To analyze the test data The sequence after sliding aggregation; It is a one-dimensional convolution.

4. The long-term multi-load forecasting method based on parallelized closed continuous neural networks according to claim 1, characterized in that, The steps for obtaining the improved liquid neural network include: Obtaining closed-form approximate solutions for liquid neural networks; The hidden states of the improved liquid neural network are obtained by replacing the exponential decay term in the closed-form approximation solution with the inverse sigmoid function; the hidden state representation formula of the improved liquid neural network is as follows: ； in, and For liquid neural networks Learnable functions; It is the inverse sigmoid function; The hidden states of the improved liquid neural network; For a specific moment; Branching features at time points; For liquid neural networks The set of all trainable weights and biases within the ; It is the Hadamard product.

5. The long-term multi-load forecasting method based on parallelized closed continuous neural networks according to claim 1, characterized in that, Before the step of splicing multiple updated hidden states to obtain spliced ​​temporal features, the method further includes: Determine the processing cycle that includes multiple time points; The time branch features at the same time within multiple consecutive processing cycles with lag are input into the improved liquid neural network to obtain the hidden states updated after multiple lag times, and then weighted and summed proportionally.

6. The long-term multi-load forecasting method based on parallelized closed continuous neural networks according to claim 1, characterized in that, The steps for performing high-frequency filtering and decoding on the spliced ​​time-series features to obtain load prediction results include: The splicing time sequence features are normalized to obtain a normalized sequence; Perform a real-signal Fourier transform on the normalized sequence along the time dimension to obtain the first transform data; The first transformed data is low-pass filtered to retain low-frequency components and cut off high-frequency noise to obtain filtered data. The filtered data is upsampled based on a single linear layer, the dimension of the filtered data is expanded and interpolated to obtain the upsampled data. The upsampled data is filled into the corresponding target frequency domain dimension based on a predetermined prediction time length to obtain filled frequency domain data; Perform an inverse Fourier transform on the filled frequency domain data and adjust the amplitude to obtain the second transform data; The second transformed data is inversely normalized, and the predicted portion is extracted as the load forecast value.

7. The long-term multi-load forecasting method based on parallelized closed continuous neural networks according to claim 1, characterized in that, The steps for obtaining the input parameters include: Determine the undetermined input parameters and output parameters, and obtain the operating data of the undetermined input parameters and output parameters. The undetermined input parameters include at least one of three features: energy auxiliary characteristics, time series index characteristics, and meteorological characteristics. The load characteristics of the output parameters include electrical load, cooling load, and / or heating load. Based on the operational data, obtain the Pearson correlation coefficient between the undetermined input parameter and the output parameter; The input parameters are determined from the undetermined input parameters based on the Pearson correlation coefficient.

8. The long-term multi-load forecasting method based on parallelized closed continuous neural networks according to claim 7, characterized in that, The steps for obtaining the runtime data include: Obtain the original data of the undetermined input and output parameters; The anomaly range of the original data is determined by the quartile method, and the original data within the anomaly range is removed to obtain the first processed data. The first processed data is processed based on the global three standard deviation principle to obtain the second processed data; The second processed data is processed based on the principle of three standard deviations locally to obtain the running data.

9. The long-term multi-load forecasting method based on parallelized closed continuous neural networks according to claim 8, characterized in that, The step of determining the outlier range of the original data using the quartile method includes: Determine the first and third quartiles of the original data; The interquartile distance is determined based on the first quartile and the third quartile, and the formula for representing the interquartile distance is as follows: ； in, Interquartile distance; It is the third quartile; It is the first quartile; The lower and upper limits of normal values ​​are obtained based on the interquartile range; the formulas for representing the lower and upper limits of normal values ​​are as follows: ； ； in, This is the lower limit threshold of the normal value; This is the upper limit threshold of the normal value; The range of abnormal values ​​is obtained based on the lower limit threshold and the upper limit threshold of normal values; The process of processing the first processed data based on the global three-standard-deviation principle to obtain the second processed data includes: A first normal range is obtained based on the mean and standard deviation of the first processed data; the formula for representing the first normal range is as follows: ； in, This is within the first normal range; For the first The first processing data of each parameter The mean, For a moment, The total number of moments; For the first The first processing data of each parameter Standard deviation; The outlier handling function removes positive values ​​outside the first normal range to obtain the second processed data; The process of processing the second processed data based on the principle of three local standard deviations to obtain the running data includes: Define the sliding window; Local data is obtained by sliding the sliding window over the second processed data. A second normal range is obtained based on the mean and standard deviation of the local data. The formula for representing the second normal range is as follows: ； ； ； ； ； in, This is the upper limit threshold of the second normal range; This is the lower limit threshold of the second normal range; For sliding windows Inner The second processing data of each parameter The mean; For sliding windows Inner The second processing data of each parameter Standard deviation; For sliding windows Internal second processing data A set of time indexes; For time indexing; For adjustment factors; Remove data from the second processed data that falls outside the second normal range, and obtain the running data.

10. A long-term multi-load forecasting system based on a parallelized closed continuous neural network, characterized in that, include: Model acquisition module (10) is used to acquire a prediction model based on a multi-period parallel closed neural network. The input parameters of the prediction model include at least one of three features: energy auxiliary features, time series index features and meteorological features. The output parameters include load features. A load forecasting module (20) is used to process the detection data of the input parameters through the forecasting model to obtain the forecasting results of the output parameters; the forecasting model includes: The multi-period sparse feature fusion submodule (21) is used to perform autocorrelation analysis on the detection data to obtain multiple downsampling periods; to downsample the detection data based on the downsampling periods to obtain subsequences; and to fuse the multiple subsequences corresponding to the multiple downsampling periods after weighted summation to obtain the fusion feature of multiple periodic time-series features. Parallel closed continuous neural network submodule (22), the parallel closed continuous neural network submodule (22) is used to divide the fusion features based on time and obtain multiple time branch features; input the multiple time branch features in parallel into the improved liquid neural network to obtain multiple updated hidden states; and concatenate the multiple updated hidden states to obtain concatenated temporal features; Frequency domain filtering submodule (23) is used to perform high-frequency filtering and decoding on the spliced ​​timing features to obtain the prediction results of the output parameters.

Images