Method and system for predicting non-inductive load change trend based on hybrid neural network

Abstract

This invention discloses a method and system for predicting the trend of imperceptible load changes based on a hybrid neural network, applicable to the field of power system load forecasting. The method includes: constructing a multi-dimensional input tensor from preprocessed imperceptible load data across multiple application scenarios; adjusting the convolutional kernels of each layer of a convolutional neural network to a shared kernel group, and extracting local features from the multi-dimensional input tensor using the convolutional neural network to obtain a multi-dimensional time feature sequence; adjusting the memory units of a recurrent neural network based on the kernel group, and extracting time-dependent features from the multi-dimensional time feature sequence using the recurrent neural network to obtain a high-dimensional time feature sequence; extracting key features from the high-dimensional time feature sequence using an attention module to obtain a comprehensive feature vector; and performing inverse standardization on the comprehensive feature vector to obtain the predicted result of the imperceptible load change trend. This invention achieves accurate prediction of the imperceptible load change trend by extracting the time-dependent features of imperceptible loads across multiple application scenarios.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and in particular to a method and system for predicting load change trends without sensory input based on a hybrid neural network. Background Technology

[0002] With the continuous construction of power grid systems, non-invisible loads from various application scenarios (civilian, industrial, commercial, and new infrastructure, etc.) are constantly being connected to the power grid system. These non-invisible loads from multiple application scenarios are characterized by diversity and uncertainty. Existing methods that rely on empirical statistics and linear models to predict the changing trends of non-invisible loads often fail to accurately capture the characteristics of non-invisible load changes over time in complex multi-application scenarios, resulting in low prediction accuracy for existing methods. Summary of the Invention

[0003] This invention provides a method and system for predicting the trend of imperceptible load changes based on a hybrid neural network, in order to solve the technical problem of low accuracy in predicting the trend of imperceptible load changes in multiple application scenarios, thereby improving the accuracy of prediction.

[0004] To address the aforementioned technical problems, embodiments of the present invention provide a method for predicting imperceptible load change trends based on a hybrid neural network. The hybrid neural network comprises a convolutional neural network, a recurrent neural network, and an attention module. The method includes: Noise reduction and time-dimensional unification processing are performed on the non-inductive load data of the target power grid under multiple application scenarios to obtain preprocessed data; The preprocessed data is constructed into a multidimensional input tensor for multiple application scenarios; The convolutional kernels of each layer of the convolutional neural network are adjusted to a group of convolutional kernels shared by multiple application scenarios. Based on the adjusted convolutional neural network, local feature extraction is performed on the multidimensional input tensor to obtain a multidimensional time feature sequence. Based on the convolutional kernel group, the memory units of each layer of the recurrent neural network are adjusted, and the time-dependent features of the multidimensional time feature sequence are extracted based on the adjusted recurrent neural network to obtain a high-dimensional time feature sequence. Based on the attention module, key features are extracted from the high-dimensional time feature sequence to obtain a comprehensive feature vector; The comprehensive feature vector is inversely normalized to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

[0005] As one preferred embodiment, adjusting the convolutional kernels of each layer of the convolutional neural network to a group of convolutional kernels shared across multiple application scenarios, and performing local feature extraction on the multidimensional input tensor based on the adjusted convolutional neural network to obtain a multidimensional time feature sequence includes: Determine the prediction time scale of the multidimensional input tensor; Based on the predicted time scale, the convolutional kernels of each layer of the convolutional neural network are adjusted to obtain a convolutional kernel group shared by multiple application scenarios; Based on the adjusted convolutional kernels and filters of each layer of the convolutional neural network, local feature extraction is performed on the multidimensional input tensor to obtain multi-timescale variation features. The multi-timescale variation features are pooled to obtain a multi-dimensional time feature sequence.

[0006] As one preferred embodiment, the step of adjusting the memory units of each layer of the recurrent neural network based on the convolutional kernel group, and extracting time-dependent features from the multidimensional time feature sequence based on the adjusted recurrent neural network to obtain a high-dimensional time feature sequence includes: Based on the time step and feature dimension of the multidimensional temporal feature sequence, the number of memory units in each layer of the recurrent neural network is adjusted; the time step and feature dimension of the multidimensional temporal feature sequence are determined based on the convolutional kernel group. The multidimensional time feature sequence is input into the adjusted recurrent neural network according to the time step described above. Based on the gating strategy of each memory unit, the temporal dependencies in the multidimensional time feature sequence are extracted to obtain a high-dimensional time feature sequence.

[0007] As one preferred embodiment, the step of extracting key features from the high-dimensional time feature sequence based on the attention module to obtain a comprehensive feature vector includes: Based on the attention module, attention weights are calculated for the high-dimensional time features at each time step to obtain the attention weights of the high-dimensional time features at each time step. The attention weights of the high-dimensional temporal features at each time step are weighted and summed to obtain the global temporal feature sequence; The high-dimensional time feature sequence and the global time feature sequence are concatenated to obtain a comprehensive feature vector.

[0008] As one preferred embodiment, the hybrid neural network further includes a pooling layer and a fully connected layer; the step of performing inverse normalization on the comprehensive feature vector to obtain the prediction results of the imperceptible load change trend of the target power grid under various application scenarios includes: The pooling layer is used to downsample the comprehensive feature vector and the multidimensional time feature sequence. The output of the pooling layer is mapped to the output layer corresponding to each application scenario through the fully connected layer to obtain the preliminary prediction results for each application scenario. The preliminary prediction results are inversely standardized to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

[0009] Another embodiment of the present invention provides a non-sensory load change trend prediction system based on a hybrid neural network, comprising: The data preprocessing module is used to perform noise reduction and time dimension unification processing on the non-inductive load data of the target power grid under multiple application scenarios to obtain preprocessed data. A multidimensional input tensor construction module is used to construct the preprocessed data into multidimensional input tensors for multiple application scenarios; The local feature extraction module is used to adjust the convolutional kernels of each layer of the convolutional neural network into a group of convolutional kernels shared by multiple application scenarios, and to extract local features from the multidimensional input tensor based on the adjusted convolutional neural network to obtain a multidimensional time feature sequence. The temporal dependency feature extraction module is used to adjust the memory units of each layer of the recurrent neural network based on the convolutional kernel group, and to extract temporal dependency features from the multidimensional temporal feature sequence based on the adjusted recurrent neural network to obtain a high-dimensional temporal feature sequence. The key feature extraction module is used to extract key features from the high-dimensional time feature sequence based on the attention module to obtain a comprehensive feature vector. The inverse standardization processing module is used to perform inverse standardization processing on the comprehensive feature vector to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

[0010] As one preferred embodiment, the local feature extraction module includes: A prediction time scale determination unit is used to determine the prediction time scale of the multidimensional input tensor; A kernel sharing unit is used to adjust the kernels of each layer of the convolutional neural network based on the prediction time scale to obtain a kernel group shared by multiple application scenarios. The local feature extraction unit is used to extract local features from the multidimensional input tensor based on the adjusted convolutional kernels and filters of each layer of the convolutional neural network, so as to obtain multi-timescale variation features. The pooling processing unit is used to perform pooling processing on the multi-timescale variation features to obtain a multi-dimensional time feature sequence.

[0011] As one preferred embodiment, the time-dependent feature extraction module includes: A memory unit adjustment unit is used to adjust the number of memory units in each layer of the recurrent neural network based on the time step and feature dimension of the multidimensional time feature sequence; the time step and feature dimension of the multidimensional time feature sequence are determined based on the convolutional kernel group; The sequential input unit is used to input the multidimensional time feature sequence into the adjusted recurrent neural network according to the time step. A high-dimensional time feature sequence extraction unit is used to extract features of the temporal dependencies in the multi-dimensional time feature sequence based on the gating strategy of each memory unit, so as to obtain a high-dimensional time feature sequence.

[0012] As one preferred embodiment, the key feature extraction module includes: The attention weight calculation unit is used to calculate the attention weight of the high-dimensional time features at each time step based on the attention module, so as to obtain the attention weight of the high-dimensional time features at each time step. The weighted summation unit is used to sum the attention weights of the high-dimensional temporal features at each time step to obtain the global temporal feature sequence; The feature sequence concatenation unit is used to concatenate the high-dimensional time feature sequence and the global time feature sequence to obtain a comprehensive feature vector.

[0013] As one preferred embodiment, the inverse normalization processing module includes: The downsampling unit is used to downsample the comprehensive feature vector and the multidimensional time feature sequence through a pooling layer; The output mapping unit is used to map the output of the pooling layer to the output layer corresponding to each application scenario through the fully connected layer, so as to obtain the preliminary prediction results of each application scenario. The inverse standardization processing unit is used to perform inverse standardization processing on the preliminary prediction results to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

[0014] This invention provides a method and system for predicting the trend of imperceptible load changes based on a hybrid neural network. By collecting imperceptible load data of a target power grid under multiple application scenarios, a multi-dimensional input tensor for each scenario is constructed. This multi-dimensional input tensor is then input into a hybrid neural network. The hybrid neural network provided in this invention is based on the changes in imperceptible load under multiple application scenarios, combining convolutional neural networks, recurrent neural networks, and attention modules. The hybrid neural network captures the time-dependent features and differentiated response patterns of imperceptible load data across multiple application scenarios. While sharing convolutional kernels, it extracts common patterns and differential features from the imperceptible load data across application scenarios. By performing inverse standardization on the comprehensive feature vector extracted by the hybrid neural network, the predicted trend of imperceptible load changes in the target power grid under each application scenario is obtained, achieving accurate prediction of imperceptible load change trends across multiple application scenarios. Attached Figure Description

[0015] Figure 1 This is one of the flowcharts of the method for predicting the trend of imperceptible load change based on a hybrid neural network provided by the present invention; Figure 2 This is the second flowchart of the method for predicting the trend of imperceptible load change based on a hybrid neural network provided by the present invention. Figure 3 This is a schematic diagram of the structure of the sensorless load change trend prediction system based on hybrid neural networks provided by the present invention.

[0016] Figure label: Among them, 301 is the data preprocessing module; 302 is the multidimensional input tensor construction module; 303 is the local feature extraction module; 304 is the temporal dependent feature extraction module; 305 is the key feature extraction module; and 306 is the inverse normalization processing module. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0019] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0020] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the imperceptible load change trend prediction method based on a hybrid neural network provided by the present invention. The hybrid neural network provided by the present invention includes a convolutional neural network, a recurrent neural network, and an attention module; as shown below. Figure 1 As shown, this embodiment includes steps 100 to 600, and the specific steps are as follows: Step 100: Perform noise reduction and time dimension unification processing on the non-inductive load data of the target power grid under multiple application scenarios to obtain preprocessed data; Specifically, in this embodiment, the sensorless load data of the target power grid under multiple application scenarios is represented in a unified mathematical form, and the set of application scenarios is defined as follows: ,in, Representing a residential scene, Indicates an industrial scenario. Indicates a business scenario. Representing new infrastructure scenarios For each type of application scenario, there is a corresponding set of devices. Application scenarios medium equipment At any moment The load power is expressed as The value is determined by the curve of installed capacity versus load utilization, as shown in Formula 1, where... For the installed capacity of the equipment, For the equipment at all times The load utilization curve.

[0022] (1) The data on the imperceptible load in residential scenarios mainly comes from loads such as air conditioners, water heaters, and heating equipment. The collected data mainly includes installed capacity, load curves per unit time, the relationship between temperature and power response, and user comfort constraints. Industrial scenarios mainly involve equipment related to production processes, machining loads, and smelting loads. The collected data mainly includes the rated capacity of equipment, load curves under production shifts, interruptibility ratios, and production continuity constraints. Commercial scenarios mainly involve central air conditioning, lighting systems, and elevators. The collected data mainly includes load distribution curves per unit time, operating time patterns, and adjustment ranges. New infrastructure scenarios mainly involve data centers, 5G base stations, electric vehicle charging piles, and energy storage systems. The collected data mainly includes installed capacity, task scheduling characteristics, charging and discharging power curves, and response speed.

[0023] The non-invisible load data of the target power grid under multiple application scenarios is preprocessed to eliminate noise and dimensional differences, and to ensure consistency in the time dimension, resulting in preprocessed data. In this embodiment, the non-invisible load data of the target power grid under multiple application scenarios comes from the power grid electricity consumption information collection system, building energy consumption management system, industrial production scheduling system, and the operation monitoring platform of new energy facilities, ensuring the comprehensiveness and representativeness of the data.

[0024] Step 200: Construct the preprocessed data into a multi-dimensional input tensor for multiple application scenarios; After obtaining the preprocessed data, the feature data of the non-invisible load of various application scenarios are constructed into an input dataset. This input dataset contains both static features (e.g., installed capacity, response capability, and operational constraints) and dynamic features (e.g., load curve per unit time and the impact of environmental conditions on the load), and is represented in the form of a multi-dimensional vector, namely a multi-dimensional input tensor for multiple application scenarios.

[0025] Step 300: Adjust the convolutional kernels of each layer of the convolutional neural network to a group of convolutional kernels shared by multiple application scenarios, and extract local features from the multidimensional input tensor based on the adjusted convolutional neural network to obtain a multidimensional time feature sequence; Specifically, the hybrid neural network provided in this embodiment includes a convolutional neural network. The convolutional layers of the convolutional neural network adopt a shared backbone structure, uniformly extracting features from multi-dimensional input tensors across multiple application scenarios. This avoids setting separate convolutional kernels for each scenario, thereby maintaining consistency across scenarios at the parameter level and providing a unified feature representation for subsequent cross-scenario commonalities. The activation function for the convolutional kernel can be ReLU to enhance the nonlinear expressive power of the hybrid neural network and improve the sparsity of feature extraction.

[0026] The number of convolutional kernels can be determined by considering the complexity of the non-intrusive load data and the accuracy requirements of the prediction task. Based on the shared convolutional kernels, this embodiment also proposes a cross-scene common law learning mechanism, which achieves unified extraction of temporal patterns, operational constraints, and energy response features across different scenarios through shared convolutional kernels, resulting in a multi-dimensional temporal feature sequence.

[0027] Step 400: Adjust the memory units of each layer of the recurrent neural network based on the convolutional kernel group, and extract time-dependent features from the multidimensional time feature sequence based on the adjusted recurrent neural network to obtain a high-dimensional time feature sequence; Specifically, the hybrid neural network provided in this embodiment also includes a recurrent neural network. After the convolutional neural network completes local feature extraction, this invention further introduces recurrent neural units, such as a Long Short-Term Memory (LSTM) network, to capture the long-term dependencies of insensitive load data over time. The multidimensional feature sequence output by the convolutional neural network is input into the LSTM. The LSTM learns the periodicity and long-term trend of the load curve through recursive operations at time steps, thereby supplementing the modeling capability of the convolutional neural network in the time dimension and improving the accuracy of characterizing long-term fluctuations and delayed responses. The multidimensional time feature sequence is input into the LSTM to obtain the high-dimensional time features output by the LSTM. The number of memory units in each layer of the recurrent neural network is related to the multidimensional time feature sequence, which in turn is associated with the shared convolutional kernel group. That is, the number of memory units in each layer of the recurrent neural network is adjusted based on the convolutional kernel group.

[0028] Step 500: Extract key features from the high-dimensional time feature sequence based on the attention module to obtain a comprehensive feature vector; Specifically, the hybrid neural network provided in this embodiment also includes an attention module. Based on the high-dimensional temporal features output by the LSTM, an attention module is introduced to adaptively assign importance weights to different time segments and feature dimensions. The attention module can automatically identify key time periods, key devices, and key scene features. Under multi-scene input conditions, it highlights representative temporal patterns and differentiated response rules, thereby achieving feature weighting and dynamic fusion, enhancing the interpretability and robustness of the hybrid neural network in multi-scene prediction tasks. The comprehensive feature vector processed by the LSTM and attention module is then input together with the output of the convolutional neural network into subsequent pooling and fully connected layers, achieving joint learning of time-dependent features, local features, and static features. This results in a comprehensive feature vector that integrates the local features of the convolutional layers, the long-term dependency features of the LSTM, and the key feature weight information of the attention module.

[0029] Step 600: Perform inverse normalization on the comprehensive feature vector to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

[0030] Specifically, the normalized predicted values ​​output by the hybrid neural network need to be converted back to their original dimensions. Using the normalization parameters saved during the training phase, each feature is independently inversely transformed to ensure the predicted values ​​remain within a reasonable range. The hybrid neural network outputs predicted power sequences for devices at different times under various scenarios. The prediction results for each scenario are generated by its independent output layer, thus maintaining differentiated prediction capabilities while sharing features. By continuously expanding the prediction results along the time dimension, the changing trends of loads in multiple scenarios can be obtained. Further aggregation and comparative analysis along the scenario dimension can reveal the overall changing trends and clustering characteristics of loads across multiple scenarios. Furthermore, combining error analysis and trend judgment of the prediction results can provide strong support for long-term power grid planning and the quantitative development of the potential for imperceptible load regulation.

[0031] This embodiment collects non-impact load data of the target power grid under multiple application scenarios, constructs a multi-dimensional input tensor for each scenario, and inputs this multi-dimensional input tensor into a hybrid neural network. The hybrid neural network provided by this invention is based on the changes in non-impact load under multiple application scenarios, combined with convolutional neural networks, recurrent neural networks, and attention modules. The hybrid neural network captures the time-dependent features and differentiated response patterns of non-impact load data across multiple application scenarios. While sharing convolutional kernels, it extracts common patterns and differential features from the non-impact load data across application scenarios. By performing inverse standardization on the comprehensive feature vector extracted by the hybrid neural network, the prediction results of the non-impact load change trends of the target power grid under each application scenario are obtained, achieving accurate prediction of non-impact load change trends under multiple application scenarios.

[0032] See Figure 2 , Figure 2 This is a flowchart illustrating another embodiment of the imperceptible load change trend prediction method based on hybrid neural networks provided by the present invention, as shown below. Figure 2 As shown, this embodiment includes steps 310 to 340, and the specific steps are as follows: Step 310: Determine the prediction time scale of the multidimensional input tensor; Step 320: Adjust the convolution kernels of each layer of the convolutional neural network based on the predicted time scale to obtain a convolution kernel group shared by multiple application scenarios; Step 330: Based on the adjusted convolutional kernels and filters of each layer of the convolutional neural network, local feature extraction is performed on the multidimensional input tensor to obtain multi-timescale variation features; Step 340: Perform pooling processing on the multi-timescale variation features to obtain a multi-dimensional time feature sequence.

[0033] Specifically, when the sensorless load data to be predicted is input into the constructed and trained hybrid neural network, the hybrid neural network first receives a multi-dimensional time series data matrix containing the following key dimensions: batch size: the number of samples processed at one time, such as 32 or 64; time step: the length of the sensorless load data to be predicted, such as 168 hours of load data; feature dimension: the number of features at each time point, including load value, temperature, humidity, and time features, etc. For example, the input data is (32, 168, 15), representing 32 samples, each sample containing 168 historical time points, and 15 features at each time point. After the sensorless load data to be predicted is preprocessed, it is converted into a standardized three-dimensional input tensor, ready to be input into the hybrid neural network.

[0034] Taking a three-layer convolutional neural network as an example, the processing procedure of the convolutional neural network is as follows: The three-dimensional input tensor first enters the first one-dimensional convolution layer. Input: [32, 168, 15] Convolution kernel: size 3, 64 filters; Operation: sliding window convolution calculation; Output: [32, 168, 64]. The convolution kernel slides in the time dimension, focusing on 3 consecutive time points each time; each filter learns a specific local load pattern, and 64 filters generate 64 different feature maps. The ReLU activation function is used to introduce nonlinearity; after batch normalization and max pooling, the output of the first one-dimensional convolution layer [32, 84, 64] is obtained.

[0035] The second and third convolutional layers have similar processing steps, but different outputs. The second convolutional layer takes input [32, 84, 64] as input; a kernel of size 5 with 128 filters, expanding the processing range to capture daily trends; and outputs [32, 84, 128]; after pooling, it yields [32, 42, 128]. The third convolutional layer takes input [32, 42, 128] as input; a kernel of size 7 with 256 filters, focusing on a longer time range to identify weekly trends; and outputs [32, 42, 256]; after pooling, it yields [32, 21, 256]. This three-layer convolutional neural network compresses the original 168 time points to 21 high-level feature points, each containing a 256-dimensional feature representation. The second convolutional layer's kernel captures the daily trend, while the third convolutional layer's kernel captures the weekly trend. In other words, the size of the convolutional kernels in the shared kernel group (each layer of the convolutional neural network) is related to the prediction time scale; the larger the prediction time scale, the larger the corresponding kernel. The final multidimensional time feature sequence output by the three-layer convolutional neural network is [32, 21, 256], which also serves as the input to the subsequent recurrent neural network.

[0036] This embodiment addresses the temporal and scenario-specific characteristics of non-sensory load data by using a convolutional neural network with shared kernels to effectively capture local features of non-sensory load data at different time scales, providing a unified feature representation foundation for subsequent cross-scenario common pattern learning.

[0037] In another embodiment of the imperceptible load change trend prediction method based on hybrid neural networks provided by the present invention, step 400 specifically includes: Step 410: Adjust the number of memory units in each layer of the recurrent neural network based on the time step and feature dimension of the multidimensional time feature sequence; the time step and feature dimension of the multidimensional time feature sequence are determined based on the convolutional kernel group; Step 420: Input the multidimensional time feature sequence into the adjusted recurrent neural network according to the time step; Step 430: Based on the gating strategy of each memory unit, extract the temporal dependencies in the multidimensional time feature sequence to obtain a high-dimensional time feature sequence.

[0038] Specifically, before inputting the multidimensional time feature sequence [32,21,256] into the recurrent neural network for processing, the number of memory units in each layer of the recurrent neural network is adjusted based on the time step and feature dimension of the multidimensional time feature sequence. For example, if the time step of the multidimensional time feature sequence [32,21,256] is 21 and the feature dimension is 256, the number of memory units in the first layer of the recurrent neural network can be adjusted to 128, and the number of memory units in the second layer of the recurrent neural network can be adjusted to 64.

[0039] The multidimensional temporal feature sequence [32, 21, 256] is input into the first layer of a recurrent neural network with 128 memory units. The forward LSTM processing is as follows: at time step t=1, initialize the memory state and hidden state to zero; calculate the forget gate, input gate, and output gate; update the cell state; calculate the hidden state; repeat until time step t=21. The backward LSTM processing is as follows: at time step t=21, initialize the state to zero; process the sequence in reverse temporal order; learn the dependencies between the future and the past. The bidirectional fusion processing is as follows: the forward and backward hidden states at each time step are concatenated to form a 256-dimensional (128+128) hidden state, and the first layer of the recurrent neural network outputs [32, 21, 256].

[0040] The output [32, 21, 256] from the first layer of the recurrent neural network (RNN) is randomly deactivated and then fed into the second layer of the RNN with 64 memory units. The processing of the second layer is similar to that of the first layer, but the number of memory units is halved, forming a feature pyramid. The second layer learns more abstract temporal features, and random deactivation is applied again to prevent overfitting. The output of the second layer of the RNN is [32, 21, 128]. The 128-dimensional feature sequence output by the LSTM (i.e., the high-dimensional temporal feature sequence in this embodiment) contains rich temporal dependency information.

[0041] The LSTM provided in this embodiment learns the periodicity and long-term trend of the load curve through recursive operations at time steps, thereby supplementing the analysis and prediction capabilities of the convolutional neural network in the time dimension and improving the prediction accuracy of the hybrid neural network for long-term fluctuations and delayed responses.

[0042] In another embodiment of the imperceptible load change trend prediction method based on hybrid neural networks provided by the present invention, step 500 specifically includes: Step 510: Based on the attention module, calculate the attention weights of the high-dimensional time features at each time step to obtain the attention weights of the high-dimensional time features at each time step. Step 520: Sum the attention weights of the high-dimensional temporal features at each time step to obtain the global temporal feature sequence; Step 530: Concatenate the high-dimensional time feature sequence and the global time feature sequence to obtain a comprehensive feature vector.

[0043] Specifically, the attention module's processing is as follows: 1. Multi-head attention decomposition: The high-dimensional temporal feature sequence output by the LSTM is copied to multiple (e.g., 8) parallel attention heads, each processed as [32, 21, 16]. Each attention head performs the following calculations separately: 1.1 Linear projection: Query vector, Key vector, Value vector, ,in, It is the output of LSTM. These are learnable parameters. 1.2 Attention Score Calculation: The correlation between all time point pairs is calculated using the Q vector, K vector, and V vector mentioned above. Time points with high similarity receive high scores. 1.3 Weighted Fusion: The attention scores are used to weight and sum the Value vectors. High-weight time point features are highlighted, and the weighted features of each attention head are output [32,16].

[0044] 2. Multi-head Feature Fusion: The outputs of eight attention heads are concatenated and fused. 2.1. Head Concatenation: The outputs of the eight [32, 16] heads are concatenated into [32, 128]. 2.2. Linear Transformation: A learnable weight matrix is ​​used... 2.3 Projection; 2.4 Residual connection; 2.5 Feedforward network: two fully connected layers with ReLU activation in the middle; 2.6 Final output [32,128], the time dimension is compressed.

[0045] This embodiment automatically identifies key time periods, key equipment, and key scene features through an attention module. Under multi-scene input conditions, it highlights representative temporal features and differentiated response patterns, thereby achieving feature weighting and dynamic fusion, and enhancing the interpretability and robustness of the hybrid neural network in multi-scene prediction tasks.

[0046] In another embodiment of the imperceptible load change trend prediction method based on hybrid neural networks provided by the present invention, step 600 specifically includes: Step 610: Downsample the comprehensive feature vector and the multidimensional time feature sequence through the pooling layer; Step 620: Through the fully connected layer, map the output of the pooling layer to the output layer corresponding to each application scenario to obtain the preliminary prediction results for each application scenario; Step 630: Perform inverse standardization on the preliminary prediction results to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

[0047] Specifically, the integrated feature vector processed by the recurrent neural network and attention module is input together with the output of the convolutional neural network into subsequent pooling and fully connected layers, achieving joint learning of time-dependent features, local features, and static features. For the n future time points to be predicted, the hybrid neural network establishes n parallel output branches. Assuming the prediction is for the load over the next 24 hours (n=24), each output branch is independent but shares the preceding feature extraction layer. The pooling layer downsamples the integrated feature vector and multi-dimensional time feature sequence, including max pooling and average pooling. The fully connected layer maps the output of the pooling layer to the output layer corresponding to each application scenario, obtaining preliminary prediction results for each application scenario.

[0048] Hybrid neural networks output scene during prediction. ,equipment ,time The predicted power output from individual devices alone cannot directly reflect the overall energy consumption characteristics of a specific scenario. Since a single scenario often includes multiple heterogeneous devices, it is necessary to sum the predicted power output of all devices in that scenario at the same time to form a scenario-level load curve. However, relying solely on the prediction results of a single scenario is insufficient to comprehensively reflect the overall imperceptible load change trend of the power grid. Therefore, further aggregation across multiple scenario dimensions is required, unifying and summing the load curves from residential, industrial, commercial, and new infrastructure scenarios to obtain a multi-scenario load prediction curve, which represents the predicted imperceptible load change trend of the target power grid in each application scenario in this embodiment.

[0049] To reflect the changing trend of scenario load over time, appropriate time segments are selected, and the load levels of future periods are compared with those of the baseline period. By calculating the difference and ratio between the load levels of future periods and the baseline period, the magnitude and relative rate of load change can be quantified. Furthermore, based on the changing trends of load curves from multiple scenarios, it can be determined whether there are synchronous or coupled relationships between the load curves of different scenarios during their changes.

[0050] This embodiment configures an independent output branch for each scenario at the output end, thereby generating prediction results for the corresponding scenario based on shared features, achieving unified analysis and prediction of cross-scenario commonalities and scenario differences.

[0051] Based on the description of the application process of the hybrid neural network above, its training process can be derived, including the construction of training data and the construction process of each part of the hybrid neural network.

[0052] The following describes the sensorless load change trend prediction system based on hybrid neural networks provided by the present invention. The sensorless load change trend prediction system based on hybrid neural networks described below can be referred to in correspondence with the sensorless load change trend prediction method based on hybrid neural networks described above.

[0053] Please refer to Figure 3 The present invention also provides a system for predicting the trend of imperceptible load changes based on a hybrid neural network, comprising: The data preprocessing module 301 is used to perform noise reduction and time dimension unification processing on the non-inductive load data of the target power grid under multiple application scenarios to obtain preprocessed data. The multidimensional input tensor construction module 302 is used to construct the preprocessed data into a multidimensional input tensor for multiple application scenarios. The local feature extraction module 303 is used to adjust the convolutional kernels of each layer of the convolutional neural network into a group of convolutional kernels shared by multiple application scenarios, and to extract local features from the multidimensional input tensor based on the adjusted convolutional neural network to obtain a multidimensional time feature sequence. The temporal dependency feature extraction module 304 is used to adjust the memory units of each layer of the recurrent neural network based on the convolutional kernel group, and to extract temporal dependency features from the multidimensional temporal feature sequence based on the adjusted recurrent neural network to obtain a high-dimensional temporal feature sequence. Key feature extraction module 305 is used to extract key features from the high-dimensional time feature sequence based on the attention module to obtain a comprehensive feature vector; The inverse standardization processing module 306 is used to perform inverse standardization processing on the comprehensive feature vector to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

[0054] Optionally, the local feature extraction module includes: A prediction time scale determination unit is used to determine the prediction time scale of the multidimensional input tensor; A kernel sharing unit is used to adjust the kernels of each layer of the convolutional neural network based on the prediction time scale to obtain a kernel group shared by multiple application scenarios. The local feature extraction unit is used to extract local features from the multidimensional input tensor based on the adjusted convolutional kernels and filters of each layer of the convolutional neural network, so as to obtain multi-timescale variation features. The pooling processing unit is used to perform pooling processing on the multi-timescale variation features to obtain a multi-dimensional time feature sequence.

[0055] Optionally, the time-dependent feature extraction module includes: A memory unit adjustment unit is used to adjust the number of memory units in each layer of the recurrent neural network based on the time step and feature dimension of the multidimensional time feature sequence; the time step and feature dimension of the multidimensional time feature sequence are determined based on the convolutional kernel group; The sequential input unit is used to input the multidimensional time feature sequence into the adjusted recurrent neural network according to the time step. A high-dimensional time feature sequence extraction unit is used to extract features of the temporal dependencies in the multi-dimensional time feature sequence based on the gating strategy of each memory unit, so as to obtain a high-dimensional time feature sequence.

[0056] Optionally, the key feature extraction module includes: The attention weight calculation unit is used to calculate the attention weight of the high-dimensional time features at each time step based on the attention module, so as to obtain the attention weight of the high-dimensional time features at each time step. The weighted summation unit is used to sum the attention weights of the high-dimensional temporal features at each time step to obtain the global temporal feature sequence; The feature sequence concatenation unit is used to concatenate the high-dimensional time feature sequence and the global time feature sequence to obtain a comprehensive feature vector.

[0057] Optionally, the inverse normalization processing module includes: The downsampling unit is used to downsample the comprehensive feature vector and the multidimensional time feature sequence through a pooling layer; The output mapping unit is used to map the output of the pooling layer to the output layer corresponding to each application scenario through the fully connected layer, so as to obtain the preliminary prediction results of each application scenario. The inverse standardization processing unit is used to perform inverse standardization processing on the preliminary prediction results to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

[0058] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for predicting the trend of imperceptible load changes based on a hybrid neural network, characterized in that, The hybrid neural network comprises a convolutional neural network, a recurrent neural network, and an attention module; the method includes: Noise reduction and time-dimensional unification processing are performed on the non-inductive load data of the target power grid under multiple application scenarios to obtain preprocessed data; The preprocessed data is constructed into a multidimensional input tensor for multiple application scenarios; The convolutional kernels of each layer of the convolutional neural network are adjusted to a group of convolutional kernels shared by multiple application scenarios. Based on the adjusted convolutional neural network, local feature extraction is performed on the multidimensional input tensor to obtain a multidimensional time feature sequence. Based on the convolutional kernel group, the memory units of each layer of the recurrent neural network are adjusted, and the time-dependent features of the multidimensional time feature sequence are extracted based on the adjusted recurrent neural network to obtain a high-dimensional time feature sequence. Based on the attention module, key features are extracted from the high-dimensional time feature sequence to obtain a comprehensive feature vector; The comprehensive feature vector is inversely normalized to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

2. The method for predicting the trend of imperceptible load changes based on a hybrid neural network as described in claim 1, characterized in that, The step of adjusting the convolutional kernels of each layer of the convolutional neural network to a group of convolutional kernels shared by multiple application scenarios, and extracting local features from the multidimensional input tensor based on the adjusted convolutional neural network to obtain a multidimensional time feature sequence includes: Determine the prediction time scale of the multidimensional input tensor; Based on the predicted time scale, the convolutional kernels of each layer of the convolutional neural network are adjusted to obtain a convolutional kernel group shared by multiple application scenarios; Based on the adjusted convolutional kernels and filters of each layer of the convolutional neural network, local feature extraction is performed on the multidimensional input tensor to obtain multi-timescale variation features. The multi-timescale variation features are pooled to obtain a multi-dimensional time feature sequence.

3. The method for predicting the trend of imperceptible load changes based on a hybrid neural network as described in claim 1, characterized in that, The step of adjusting the memory units of each layer of the recurrent neural network based on the convolutional kernel group, and extracting time-dependent features from the multidimensional time feature sequence based on the adjusted recurrent neural network to obtain a high-dimensional time feature sequence includes: Based on the time step and feature dimension of the multidimensional temporal feature sequence, the number of memory units in each layer of the recurrent neural network is adjusted; the time step and feature dimension of the multidimensional temporal feature sequence are determined based on the convolutional kernel group. The multidimensional time feature sequence is input into the adjusted recurrent neural network according to the time step described above. Based on the gating strategy of each memory unit, the temporal dependencies in the multidimensional time feature sequence are extracted to obtain a high-dimensional time feature sequence.

4. The method for predicting the trend of imperceptible load changes based on a hybrid neural network as described in claim 1, characterized in that, The key feature extraction of the high-dimensional time feature sequence based on the attention module to obtain the comprehensive feature vector includes: Based on the attention module, attention weights are calculated for the high-dimensional time features at each time step to obtain the attention weights of the high-dimensional time features at each time step. The attention weights of the high-dimensional temporal features at each time step are weighted and summed to obtain the global temporal feature sequence; The high-dimensional time feature sequence and the global time feature sequence are concatenated to obtain a comprehensive feature vector.

5. The method for predicting the trend of imperceptible load changes based on a hybrid neural network as described in claim 1, characterized in that, The hybrid neural network further includes a pooling layer and a fully connected layer; the inverse normalization of the comprehensive feature vector to obtain the prediction results of the imperceptible load change trend of the target power grid under various application scenarios includes: The pooling layer is used to downsample the comprehensive feature vector and the multidimensional time feature sequence. The output of the pooling layer is mapped to the output layer corresponding to each application scenario through the fully connected layer to obtain the preliminary prediction results for each application scenario. The preliminary prediction results are inversely standardized to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

6. A non-sensory load change trend prediction system based on hybrid neural networks, characterized in that, include: The data preprocessing module is used to perform noise reduction and time dimension unification processing on the non-inductive load data of the target power grid under multiple application scenarios to obtain preprocessed data. A multidimensional input tensor construction module is used to construct the preprocessed data into multidimensional input tensors for multiple application scenarios; The local feature extraction module is used to adjust the convolutional kernels of each layer of the convolutional neural network into a group of convolutional kernels shared by multiple application scenarios, and to extract local features from the multidimensional input tensor based on the adjusted convolutional neural network to obtain a multidimensional time feature sequence. The temporal dependency feature extraction module is used to adjust the memory units of each layer of the recurrent neural network based on the convolutional kernel group, and to extract temporal dependency features from the multidimensional temporal feature sequence based on the adjusted recurrent neural network to obtain a high-dimensional temporal feature sequence. The key feature extraction module is used to extract key features from the high-dimensional time feature sequence based on the attention module to obtain a comprehensive feature vector. The inverse standardization processing module is used to perform inverse standardization processing on the comprehensive feature vector to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

7. The sensorless load change trend prediction system based on hybrid neural networks as described in claim 6, characterized in that, The local feature extraction module includes: A prediction time scale determination unit is used to determine the prediction time scale of the multidimensional input tensor; A kernel sharing unit is used to adjust the kernels of each layer of the convolutional neural network based on the prediction time scale to obtain a kernel group shared by multiple application scenarios. The local feature extraction unit is used to extract local features from the multidimensional input tensor based on the adjusted convolutional kernels and filters of each layer of the convolutional neural network, so as to obtain multi-timescale variation features. The pooling processing unit is used to perform pooling processing on the multi-timescale variation features to obtain a multi-dimensional time feature sequence.

8. The sensorless load change trend prediction system based on hybrid neural networks as described in claim 6, characterized in that, The time-dependent feature extraction module includes: A memory unit adjustment unit is used to adjust the number of memory units in each layer of the recurrent neural network based on the time step and feature dimension of the multidimensional time feature sequence; the time step and feature dimension of the multidimensional time feature sequence are determined based on the convolutional kernel group; The sequential input unit is used to input the multidimensional time feature sequence into the adjusted recurrent neural network according to the time step. A high-dimensional time feature sequence extraction unit is used to extract features of the temporal dependencies in the multi-dimensional time feature sequence based on the gating strategy of each memory unit, so as to obtain a high-dimensional time feature sequence.

9. The sensorless load change trend prediction system based on hybrid neural networks as described in claim 6, characterized in that, The key feature extraction module includes: The attention weight calculation unit is used to calculate the attention weight of the high-dimensional time features at each time step based on the attention module, so as to obtain the attention weight of the high-dimensional time features at each time step. The weighted summation unit is used to sum the attention weights of the high-dimensional temporal features at each time step to obtain the global temporal feature sequence; The feature sequence concatenation unit is used to concatenate the high-dimensional time feature sequence and the global time feature sequence to obtain a comprehensive feature vector.

10. The sensorless load change trend prediction system based on hybrid neural networks as described in claim 6, characterized in that, The inverse normalization processing module includes: The downsampling unit is used to downsample the comprehensive feature vector and the multidimensional time feature sequence through a pooling layer; The output mapping unit is used to map the output of the pooling layer to the output layer corresponding to each application scenario through the fully connected layer, so as to obtain the preliminary prediction results of each application scenario. The inverse standardization processing unit is used to perform inverse standardization processing on the preliminary prediction results to obtain the prediction results of the non-impact load change trend of the target power grid under various application scenarios.

Images