Electric power spatio-temporal data efficient feature extraction method based on neural network

Abstract

The invention discloses a power spatio-temporal data efficient feature extraction method based on a neural network, and relates to the field of data processing. The power spatio-temporal data efficient feature extraction method based on the neural network comprises the following steps: step 1, data preparation; 2, designing a model; step 3, model fusion; 4, setting and training parameters; and step 5, evaluation and verification. According to the electric power spatio-temporal data efficient feature extraction method based on the neural network, the features of different network architectures of the CNN, the RNN and the SNN are integrated to construct a more powerful and higher-adaptability model, the advantages of different networks during processing of specific types of data are integrated, and the feature extraction efficiency of the electric power spatio-temporal data is improved. Wherein the method includes but is not limited to the capability of the CNN in image feature extraction, the advantages of the RNN in sequence data processing, and the potential of the SNN in simulation of biological neuron activity.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and specifically to an efficient feature extraction method for power spatio-temporal data based on neural networks. Background Art

[0002] Power system dispatching refers to the command, supervision, and management activities in the process of supplying electric energy provided by many power plants to a large number of users through a power transmission, transformation, distribution, and power supply network. Its main function is to ensure the safety, economy, and efficiency of power production and operation, while meeting the electricity demand of users.

[0003] Traditional load forecasting methods often rely on simple statistical models or time series analysis, and it is difficult to accurately capture the complex changes of the load. Traditional fault diagnosis methods rely on manual inspections and empirical judgments, which are time-consuming and error-prone. The optimal dispatching of power resources needs to comprehensively consider multiple factors, such as the power supply-demand relationship, transmission line capacity, substation load, etc. The amount of data in the power field is huge and complex, and traditional data processing methods are difficult to handle.

[0004] Therefore, this application proposes an efficient feature extraction method for power spatio-temporal data based on neural networks. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides an efficient feature extraction method for power spatio-temporal data based on neural networks, which solves the problems that traditional load forecasting methods often rely on simple statistical models or time series analysis and it is difficult to accurately capture the complex changes of the load. Traditional fault diagnosis methods rely on manual inspections and empirical judgments, which are time-consuming and error-prone. The optimal dispatching of power resources needs to comprehensively consider multiple factors, such as the power supply-demand relationship, transmission line capacity, substation load, etc. The amount of data in the power field is huge and complex, and traditional data processing methods are difficult to handle.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: An efficient feature extraction method for power spatio-temporal data based on neural networks, including the following steps:

[0007] Step 1, Data Preparation: It includes power load forecasting data preparation, power system fault diagnosis data preparation, and power resource optimal dispatching data preparation;

[0008] Step 2, Model Design: It includes a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), and a Spiking Neural Network (SNN);

[0009] Step 3, Model Fusion: It includes serial fusion, parallel fusion, and hybrid layer fusion;

[0010] Step 4, Parameter Setting and Training: It includes parameter setting, training process, and optimization algorithm;

[0011] Step Five, Evaluation and Verification: It includes evaluation metrics, cross-validation, and result analysis;

[0012] Integrate the characteristics of different network architectures of Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Spiking Neural Network (SNN) to build a more powerful and adaptable model, and synthesize the advantages of different networks in processing specific types of data, including but not limited to the ability of CNN in image feature extraction, the advantage of RNN in sequence data processing, and the potential of SNN in simulating biological neuron activities.

[0013] Preferably, the power load forecasting data preparation includes:

[0014] Data collection, collect historical load data, including but not limited to time series load values, dates, times, weather conditions, holiday information, and economic activity indicators, ensure the integrity and accuracy of the data, and fill in and correct missing and abnormal data;

[0015] Data cleaning, remove duplicate data, invalid data, and abnormal data, and normalize the data to improve the training efficiency and prediction accuracy of the model;

[0016] The power system fault diagnosis data preparation includes:

[0017] Data collection, collect the operation status data of power equipment, including but not limited to sensor data of current, voltage, temperature, and vibration, and collect fault record data, including but not limited to fault types, fault times, and fault location information;

[0018] Data cleaning, filter and denoise the sensor data to improve the reliability of the data, and verify and correct the fault record data to ensure the accuracy of the fault information;

[0019] The power resource optimal scheduling data preparation:

[0020] Data collection, collect the spatio-temporal distribution data of power resources, including but not limited to the output of power plants, the transmission capacity of transmission lines, and the load of substations, and collect market demand data, including but not limited to user electricity demand and electricity price information.

[0021] Data cleaning, verify and correct the spatio-temporal distribution data of power resources to ensure the accuracy and consistency of the data, and clean and sort the market demand data to remove invalid and abnormal data.

[0022] Preferably, the Convolutional Neural Network (CNN) is used to extract the spatial features of data, and multiple convolutional layers can be designed, with an activation function and a pooling layer following each convolutional layer;

[0023] The recurrent neural network (RNN) is used to capture the temporal dependencies in the data. The spatial feature sequence extracted by the CNN is used as the input of the RNN, and the output of the RNN can be further used for tasks such as classification, regression, and time series prediction.

[0024] The spiking neural network (SNN) is introduced to simulate the spiking behavior of biological neurons. The SNN captures the dynamic characteristics of spatio-temporal data through its unique spike coding and synaptic transmission mechanisms.

[0025] Preferably, the serial fusion is specifically as follows: First, the CNN is used to extract spatial features, and then the feature sequence is input into the RNN to capture the temporal dependencies.

[0026] The parallel fusion is specifically as follows: The outputs of the CNN and the SNN are processed in parallel, and then combined with the RNN for processing in the time dimension.

[0027] The hybrid layer fusion is specifically as follows: A hybrid layer is introduced into the model to more tightly integrate the characteristics of the CNN, RNN, and SNN.

[0028] Preferably, the parameter setting is specifically as follows: According to the characteristics of the dataset and the model structure, parameters such as the learning rate, batch size, and number of iterations are set.

[0029] The training process is specifically as follows: The backpropagation algorithm is used to train the CNN and RNN models, and the spike-timing-dependent plasticity (STDP) algorithm is used to train the SNN model.

[0030] The optimization algorithm is specifically as follows: A suitable optimizer (such as Adam, SGD) is selected to accelerate the training process and reduce overfitting.

[0031] Preferably, the evaluation metric is specifically as follows: Appropriate evaluation metrics are selected according to the specific task, including but not limited to accuracy, recall, F1-score, mean squared error (MSE).

[0032] The cross-validation is specifically as follows: The K-fold cross-validation method is adopted to evaluate the generalization ability of the model.

[0033] The result analysis is specifically as follows: Visual analysis is performed on the features extracted by the model to verify whether it can accurately reflect the characteristics and laws of spatio-temporal data.

[0034] Preferably, the hybrid layer design includes the hybrid layer architecture, training and optimization, and implementation and evaluation.

[0035] The hybrid layer architecture includes: a feature extraction layer, a sequence processing layer, and a spiking neural network layer.

[0036] The training and optimization include: a hybrid loss function, backpropagation and BPTT, and hyperparameter tuning.

[0037] The implementation and evaluation include:

[0038] Implement, use deep learning frameworks (such as TensorFlow, PyTorch) to implement a hybrid layer architecture;

[0039] Evaluate, conduct experiments on standard datasets, evaluate the performance of the model, and compare it with models that only use CNN, RNN, and SNN.

[0040] Preferably, the feature extraction layer uses the input layer to receive raw data, such as images and video frames, extracts spatial hierarchical features through the CNN layer and uses multiple convolutional layers, pooling layers, and activation layers, and these features can then be used as inputs for subsequent layers;

[0041] The sequence processing layer takes the feature sequence extracted by the CNN layer as input and uses the RNN layer to process dependencies in the time dimension; this is particularly important for video processing and natural language processing tasks. And an attention mechanism is added to the RNN layer to improve the model's attention to important information.

[0042] The spiking neural network layer takes the output of the RNN layer as input and performs further processing through spiking neurons. The SNN layer brings a more sensitive and efficient response to dynamic changes by simulating the firing behavior of biological neurons. Since the RNN layer usually outputs continuous floating-point numbers, while the SNN layer needs to process spikes (i.e., discrete pulses), and a threshold activation function is used to convert the output of the RNN layer into a spike sequence processed by the SNN layer.

[0043] Preferably, the hybrid loss function is specifically: design a loss function according to the task objective (such as classification, regression);

[0044] The backpropagation and BPTT are specifically: for the RNN layer and the CNN layer parts, use traditional backpropagation algorithms and backpropagation through time algorithms for training, while the training of the SNN layer uses a training algorithm based on backpropagation of spike times;

[0045] The hyperparameter tuning is specifically: tune the hyperparameters of the number of layers, the number of neurons, and the learning rate of the CNN layer, RNN layer, and SNN layer to maximize the model performance.

[0046] Preferably, the threshold activation function is a function that converts an input signal into an output signal, and its output depends on whether the input value exceeds a specific threshold, and its expression is as follows:

[0047]

[0048] Where θ is the set threshold parameter, x is the input value, and f(x) is the output value;

[0049] By setting the threshold, the threshold activation function can limit the range of input values within a certain interval and achieve non-linear transformation of the neural network, thereby enhancing the expression ability of the model. In the input signal, only those features that exceed the threshold will be further processed by the network, which helps to ignore unimportant noises or features and improve the robustness of the model.

[0050] The present invention discloses an efficient feature extraction method for power spatio-temporal data based on a neural network, and its beneficial effects are as follows:

[0051] 1. The efficient feature extraction method for power spatio-temporal data based on a neural network collects power load prediction data, power system fault diagnosis data, and power resource optimization scheduling data, and integrates the characteristics of different network architectures of convolutional neural network (CNN), recurrent neural network (RNN), and spiking neural network (SNN). After the convolutional neural network extracts spatial features, it is input into the recurrent neural network to capture time dependence, forming a spatio-temporal feature sequence. The convolutional neural network and the spiking neural network are processed in parallel to extract spatial features and dynamic features, and combined with the recurrent neural network for time dimension processing. By introducing a hybrid layer, the characteristics of the convolutional neural network, recurrent neural network, and spiking neural network are tightly integrated to achieve more efficient feature extraction and fusion, so as to construct a more powerful and adaptable model, and synthesize the advantages of different networks in processing specific types of data; the spatio-temporal features extracted from power load prediction can accurately reflect the load change trend, improve the accuracy of load prediction, and the extracted fault features can quickly locate the fault position and type, improving the efficiency and accuracy of fault diagnosis; the spatio-temporal features extracted from power resource optimization scheduling can reflect the supply-demand relationship and transmission capacity of power resources.

[0052] 2. The efficient feature extraction method for power spatio-temporal data based on a neural network extracts the spatio-temporal features of power load by integrating convolutional neural network (CNN), recurrent neural network (RNN), and spiking neural network (SNN), significantly improving the accuracy of load prediction. This helps the power dispatching department better plan power production and transmission, reduce supply-demand imbalance and power waste. By extracting the spatio-temporal features of power resources, it provides a scientific basis for optimized scheduling. This helps to achieve reasonable allocation and efficient utilization of power resources, reduce power transmission losses and operating costs, and improve the overall efficiency of the power system.

[0053] 3. The efficient feature extraction method for power spatio-temporal data based on neural network uses deep learning technology to automatically extract the fault features of power equipment, quickly locate the fault position and type. This greatly shortens the fault diagnosis time, improves the efficiency and accuracy of fault diagnosis, helps to repair faults in time, and ensures the stable operation of the power system. By using deep learning technology, the automatic processing and efficient analysis of power data are realized. This helps to tap the potential value of power data and provides a strong guarantee for decision-making support in the power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0055] Figure 1 It is a schematic flow chart of the present invention;

[0056] Figure 2 It is a schematic flow chart of the hybrid layer design of the present invention

[0057] Figure 3 It is a schematic diagram of the threshold activation function of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described below. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0059] By providing an efficient feature extraction method for power spatio-temporal data based on neural network in the embodiments of the present application, the problem that different neural network models have different advantages and each single neural network model also has its own individual disadvantages is solved. The characteristics of different network architectures of convolutional neural network (CNN), recurrent neural network (RNN), and spiking neural network (SNN) are integrated to build a more powerful and adaptable model, and the advantages of different networks in processing specific types of data are combined, including but not limited to the ability of CNN in image feature extraction, the advantage of RNN in sequence data processing, and the potential of SNN in simulating biological neuron activities.

[0060] To better understand the above technical solutions, the following will detail the above technical solutions in conjunction with the drawings in the specification and specific embodiments.

[0061] An embodiment of the present invention discloses an efficient feature extraction method for power spatio-temporal data based on a neural network.

[0062] According to the attached Figures 1 - 3 shown, it includes the following steps:

[0063] Step 1, data preparation: It includes power load prediction data preparation, power system fault diagnosis data preparation, and power resource optimization scheduling data preparation;

[0064] Step 2, model design: It includes a convolutional neural network CNN, a recurrent neural network RNN, and a spiking neural network SNN;

[0065] Step 3, model fusion: It includes serial fusion, parallel fusion, and hybrid layer fusion;

[0066] Step 4, parameter setting and training: It includes parameter setting, training process, and optimization algorithm

[0067] Step 5, evaluation and verification: It includes evaluation metrics, cross-validation, and result analysis;

[0068] Integrate the characteristics of different network architectures of the convolutional neural network CNN, the recurrent neural network RNN, and the spiking neural network SNN to build a more powerful and adaptable model, and synthesize the advantages of different networks in processing specific types of data, including but not limited to the ability of CNN in image feature extraction, the advantage of RNN in sequence data processing, and the potential of SNN in simulating biological neuron activities.

[0069] In terms of power load prediction, by capturing the spatio-temporal dependence of load data, the accuracy of load prediction can be significantly improved. This helps power system dispatchers understand future power demands more accurately, thereby formulating more reasonable power generation and transmission plans. In terms of power system fault diagnosis, it can quickly identify abnormal states of power equipment. By extracting these features and training the model, early warnings can be issued at the initial stage of a fault, reducing the impact of the fault on the operation of the power system. In terms of power resource optimization scheduling, considering multiple factors such as the spatio-temporal distribution of power resources, market demand, and the output limits of power plants, a more reasonable scheduling strategy is formulated. This helps balance power supply and demand, reduce power losses, and improve the utilization efficiency of power resources.

[0070] Collect power load forecasting data, power system fault diagnosis data, and power resource optimal scheduling data, and integrate the characteristics of different network architectures of convolutional neural network (CNN), recurrent neural network (RNN), and spiking neural network (SNN). After the convolutional neural network extracts spatial features, it is input into the recurrent neural network to capture time dependence, forming a spatio-temporal feature sequence. The convolutional neural network and the spiking neural network are processed in parallel to extract spatial features and dynamic features, and combined with the recurrent neural network for time dimension processing. By introducing a hybrid layer, the characteristics of the convolutional neural network, recurrent neural network, and spiking neural network are tightly integrated to achieve more efficient feature extraction and fusion, so as to build a more powerful and adaptable model, and synthesize the advantages of different networks in processing specific types of data; the spatio-temporal features extracted from power load forecasting can accurately reflect the load change trend, improve the accuracy of load forecasting, and the extracted fault features can quickly locate the fault location and type, improving the efficiency and accuracy of fault diagnosis; the spatio-temporal features extracted from power resource optimal scheduling can reflect the supply-demand relationship and transmission capacity of power resources.

[0071] Further, the preparation of the power load forecasting data includes:

[0072] Data collection, collect historical load data, including but not limited to time series load values, dates, times, weather conditions, holiday information, and economic activity indicators, ensure the integrity and accuracy of the data, and fill and correct missing and abnormal data;

[0073] Data cleaning, remove duplicate data, invalid data, and abnormal data, and normalize the data to improve the training efficiency and prediction accuracy of the model;

[0074] The preparation of the power system fault diagnosis data includes:

[0075] Data collection, collect the operation status data of power equipment, including but not limited to sensor data of current, voltage, temperature, and vibration, and collect fault record data, including but not limited to fault type, fault time, and fault location information;

[0076] Data cleaning, filter and denoise the sensor data to improve the reliability of the data, and check and correct the fault record data to ensure the accuracy of the fault information;

[0077] The preparation of the power resource optimal scheduling data:

[0078] Data collection, collect the spatio-temporal distribution data of power resources, including but not limited to the output of power plants, the transmission capacity of transmission lines, and the load of substations, and collect market demand data, including but not limited to user electricity demand and electricity price information.

[0079] Data cleaning involves validating and correcting the spatio-temporal distribution data of power resources to ensure data accuracy and consistency, and cleaning and organizing the market demand data to remove invalid and abnormal data.

[0080] Furthermore, a Convolutional Neural Network (CNN) is used to extract the spatial features of data. Multiple convolutional layers can be designed, with an activation function and a pooling layer following each convolutional layer.

[0081] A Recurrent Neural Network (RNN) is used to capture the temporal dependencies in the data. The sequence of spatial features extracted by the CNN is used as the input to the RNN, and the output of the RNN can be further used for tasks such as classification, regression, and time series prediction.

[0082] A Spiking Neural Network (SNN) is introduced to simulate the spiking behavior of biological neurons. The SNN captures the dynamic characteristics of spatio-temporal data through its unique spike coding and synaptic transmission mechanisms.

[0083] Furthermore, serial fusion is specifically as follows: First, the CNN is used to extract spatial features, and then the feature sequence is input into the RNN to capture temporal dependencies.

[0084] Parallel fusion is specifically as follows: The outputs of the CNN and the SNN are processed in parallel, and then combined with the RNN for processing in the time dimension.

[0085] Hybrid layer fusion is specifically as follows: A hybrid layer is introduced into the model to more tightly integrate the characteristics of the CNN, RNN, and SNN.

[0086] Furthermore, parameter settings are specifically as follows: According to the characteristics of the dataset and the model structure, parameters such as the learning rate, batch size, and number of iterations are set.

[0087] The training process is specifically as follows: The backpropagation algorithm is used to train the CNN and RNN models, and the Spike-Timing-Dependent Plasticity (STDP) algorithm is used to train the SNN model.

[0088] Backpropagation of the CNN: The CNN extracts features from the input image through convolutional layers, activation layers, and pooling layers. During the forward propagation process, the data passes through each layer until the output layer. Backpropagation starts from the output layer, calculates the gradient of the loss function with respect to each output node, and then these gradients are propagated back to each layer through the chain rule to calculate the gradient of the loss function with respect to the parameters of that layer. Finally, the weights are updated using gradient descent (such as Adam, RMSprop).

[0089] Backpropagation in RNN: RNN processes sequential data, and its structure allows the network to share information between time steps; due to the sequential nature of RNN, the backpropagation algorithm in RNN is called Backpropagation Through Time (BPTT); BPTT unfolds the RNN through time, calculates the gradients of the loss function at each time step, and backpropagates these gradients back into the network.

[0090] STDP adjusts synaptic weights based on the relative time difference of spikes between neurons. If the presynaptic neuron fires a spike shortly before the postsynaptic neuron fires a spike, the synaptic weight increases; if the order is reversed, the synaptic weight decreases.

[0091] STDP Training in SNN: In SNN, neurons transmit information in the form of spikes rather than continuous values; when training an SNN using the STDP algorithm, first simulate the spike generation process of neurons; then, according to the time difference of spikes between the presynaptic and postsynaptic neurons, adjust the synaptic weights using the STDP rule; this process is iterated until the behavior of the network meets the expected stopping criterion.

[0092] The optimization algorithm is specifically as follows: Select a suitable optimizer (such as Adam, SGD) to accelerate the training process and reduce overfitting.

[0093] Furthermore, the evaluation metrics are specifically as follows: Select appropriate evaluation metrics according to the specific task, including but not limited to accuracy, recall, F1-score, mean squared error (MSE);

[0094] Cross-validation is specifically as follows: Adopt the method of K-fold cross-validation to evaluate the generalization ability of the model;

[0095] K-fold cross-validation reduces the risk of overfitting and gives a robust estimate of the model performance by splitting the dataset into K subsets of equal size.

[0096] Result analysis is specifically as follows: Conduct a visual analysis of the features extracted by the model to verify whether it can accurately reflect the characteristics and laws of spatio-temporal data.

[0097] Preferably, the hybrid layer design includes hybrid layer architecture, training and optimization, and implementation and evaluation;

[0098] The hybrid layer architecture includes: a feature extraction layer, a sequence processing layer, and a spiking neural network layer;

[0099] Training and optimization include: hybrid loss function, backpropagation and BPTT, and hyperparameter tuning;

[0100] Implementation and evaluation include:

[0101] Implement and use deep learning frameworks (such as TensorFlow, PyTorch) to implement a hybrid layer architecture;

[0102] Evaluate, conduct experiments on standard datasets, evaluate the performance of the model, and compare it with models that only use CNN, RNN, and SNN.

[0103] Specifically disclosed, the feature extraction layer uses the input layer to receive raw data, such as images and video frames, extracts spatial hierarchical features through the CNN layer and uses multiple convolutional layers, pooling layers, and activation layers, and these features can then be used as the input for subsequent layers;

[0104] The sequence processing layer takes the feature sequence extracted by the CNN layer as input and uses the RNN layer to process the dependencies in the time dimension; this is particularly important for video processing and natural language processing tasks. And an attention mechanism is added to the RNN layer to improve the model's attention to important information.

[0105] The spiking neural network layer takes the output of the RNN layer as input and performs further processing through spiking neurons. The SNN layer brings a more sensitive and efficient response to dynamic changes by simulating the firing behavior of biological neurons. Since the RNN layer usually outputs continuous floating-point numbers, while the SNN layer needs to process spikes (i.e., discrete pulses), a threshold activation function is used to convert the output of the RNN layer into a spike sequence processed by the SNN layer.

[0106] Specifically disclosed, the hybrid loss function is specifically: design the loss function according to the task objective (such as classification, regression);

[0107] Backpropagation and BPTT are specifically: for the RNN layer and CNN layer parts, use the traditional backpropagation algorithm and the backpropagation through time algorithm for training, while the training of the SNN layer uses the training algorithm based on backpropagation of spike times;

[0108] Hyperparameter tuning is specifically: tune the hyperparameters of the number of layers, number of neurons, and learning rate of the CNN layer, RNN layer, and SNN layer to maximize the model performance.

[0109] Furthermore, the threshold activation function is a function that converts the input signal into an output signal, and its output depends on whether the input value exceeds a specific threshold, and its expression is as follows:

[0110]

[0111] Where θ is the set threshold parameter, x is the input value, and f(x) is the output value;

[0112] By setting the threshold, the threshold activation function can limit the range of input values ​​to a certain interval and realize nonlinear transformation of the neural network, thereby enhancing the expressiveness of the model. In the input signal, only those features exceeding the threshold will be further processed by the network, which helps to ignore unimportant noise or features and improve the robustness of the model.

[0113] Through the set hybrid layer, the feature extraction layer uses the input layer to receive the original data, and extracts spatial hierarchical features through the CNN layer and multiple convolutional layers, pooling layers, and activation layers; the sequence processing layer uses the feature sequence extracted by the CNN layer as input, and uses the RNN layer to process the dependencies in the time dimension; and adds an attention mechanism to the RNN layer to improve the model's attention to important information. The spike neural network layer uses the output of the RNN layer as input and further processes it through spike neurons. The SNN layer simulates the discharge behavior of biological neurons to bring a more sensitive and efficient response to dynamic changes, and uses a threshold activation function to convert the output of the RNN layer into a spike sequence processed by the SNN layer.

[0114] The characteristics of different network architectures such as convolutional neural network (CNN), recurrent neural network (RNN) and spiking neural network (SNN) are integrated to build a more powerful and adaptable model, combining the advantages of different networks in processing specific types of data, including but not limited to the ability of CNN in image feature extraction, the advantages of RNN in sequence data processing, and the potential of SNN in simulating biological neuron activity.

[0115] By integrating convolutional neural networks (CNN), recurrent neural networks (RNN) and spiking neural networks (SNN), the temporal and spatial characteristics of power loads are extracted, significantly improving the accuracy of load forecasting. This helps power dispatching departments to better plan power production and transmission, reduce supply and demand imbalances and power waste. By extracting the temporal and spatial characteristics of power resources, a scientific basis is provided for optimizing dispatching. This helps to achieve the rational allocation and efficient use of power resources, reduce power transmission losses and operating costs, and improve the overall benefits of the power system. Using deep learning technology, the fault characteristics of power equipment are automatically extracted, and the fault location and type are quickly located. This greatly shortens the fault diagnosis time, improves the efficiency and accuracy of fault diagnosis, helps to repair faults in a timely manner, and ensures the stable operation of the power system. Using deep learning technology, automatic processing and efficient analysis of power data are achieved. This helps to tap the potential value of power data and provide strong support for decision-making support in the power industry.

[0116] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification is only to illustrate the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. An efficient feature extraction method for power spatiotemporal data based on neural network, characterized in that: The following steps are involved: Step 1: Data preparation: This includes power load forecasting data preparation, power system fault diagnosis data preparation, and power resource optimization and dispatching data preparation; Step 2: Model design: It includes convolutional neural network CNN, recurrent neural network RNN, and pulse neural network SNN; Step 3: Model fusion: It includes serial fusion, parallel fusion and mixed layer fusion; Step 4: Parameter setting and training: It includes parameter setting, training process and optimization algorithm; Step 5: Evaluation and validation: This includes evaluation indicators, cross-validation, and result analysis; The power load forecast data, power system fault diagnosis data and power resource optimization and dispatching data are collected, and the characteristics of different network architectures such as convolutional neural network (CNN), recurrent neural network (RNN) and pulse neural network (SNN) are used for integration. After the convolutional neural network extracts the spatial features, the recurrent neural network is input to capture the time dependency to form a spatiotemporal feature sequence. The convolutional neural network and the pulse neural network are processed in parallel to extract spatial features and dynamic features. The recurrent neural network is combined for time dimension processing. By introducing a hybrid layer, the characteristics of the convolutional neural network, the recurrent neural network and the pulse neural network are closely integrated to achieve more efficient feature extraction and fusion, so as to build a more powerful and adaptable model and integrate the advantages of different networks in processing specific types of data. The spatiotemporal features extracted by the power load forecast can accurately reflect the load change trend and improve the accuracy of load forecasting. The extracted fault features can quickly locate the fault location and type and improve the efficiency and accuracy of fault diagnosis. The spatiotemporal features extracted by the power resource optimization and dispatching can reflect the supply and demand relationship and transmission capacity of power resources.

2. The method for extracting efficient features of power spatiotemporal data based on neural network according to claim 1 is characterized in that: The power load forecasting data preparation includes: Data collection: collect historical load data, including but not limited to time series load values, date, time, weather conditions, holiday information and economic activity indicators, ensure the integrity and accuracy of the data, and fill in and correct missing and abnormal data; Data cleaning: removing duplicate data, invalid data and abnormal data, and normalizing the data to improve the training efficiency and prediction accuracy of the model; The power system fault diagnosis data preparation includes: Data collection: collecting operating status data of power equipment, including but not limited to sensor data of current, voltage, temperature and vibration, and collecting fault record data, including but not limited to fault type, fault time and fault location information; Data cleaning: filtering and denoising sensor data to improve data reliability, verifying and correcting fault record data to ensure the accuracy of fault information; The power resource optimization dispatching data is prepared as follows: Data collection, collecting the temporal and spatial distribution data of power resources, including but not limited to the output of power plants, the transmission capacity of transmission lines and the load of substations, and collecting market demand data, including but not limited to user electricity demand and electricity price information. Data cleaning: verify and correct the spatiotemporal distribution data of power resources to ensure the accuracy and consistency of the data, clean and organize the market demand data, and remove invalid and abnormal data.

3. The method for extracting efficient features of power spatiotemporal data based on neural network according to claim 1 is characterized in that: The convolutional neural network CNN is used to extract spatial features of data, and multiple convolutional layers can be designed, each of which is followed by an activation function and a pooling layer; The recurrent neural network RNN ​​is used to capture the temporal dependency in the data, and the spatial feature sequence extracted by CNN is used as the input of RNN. The output of RNN can be further used for classification, regression and time series prediction tasks; The spiking neural network (SNN) introduces SNN to simulate the pulse emission behavior of biological neurons. SNN captures the dynamic characteristics of spatiotemporal data through its unique pulse encoding and synaptic transmission mechanism.

4. The method for extracting efficient features of power spatiotemporal data based on neural network according to claim 1 is characterized in that: The serial fusion is specifically as follows: firstly, CNN is used to extract spatial features, and then the feature sequence is input into RNN to capture the temporal dependency; The parallel fusion is specifically: processing the outputs of CNN and SNN in parallel, and then combining RNN to process the time dimension; The hybrid layer fusion specifically includes: introducing a hybrid layer into the model to more closely integrate the characteristics of CNN, RNN and SNN.

5. The method for extracting efficient features of power spatiotemporal data based on neural network according to claim 1 is characterized in that: The parameter setting is specifically as follows: setting the parameters of learning rate, batch size, and number of iterations according to the characteristics of the data set and the model structure; The specific training process is as follows: the CNN and RNN models are trained using the back-propagation algorithm, and the SNN model is trained using the spike time-dependent plasticity STDP algorithm; The optimization algorithm is specifically: select a suitable optimizer to speed up the training process and reduce overfitting.

6. The method for efficient feature extraction of power spatiotemporal data based on neural network according to claim 1 is characterized in that: The evaluation indicators are specifically: selecting appropriate evaluation indicators according to specific tasks, including but not limited to accuracy, recall rate, F1 score, and mean square error; Cross-validation is specifically: using the K-fold cross-validation method to evaluate the generalization ability of the model; The specific results analysis includes: visual analysis of the features extracted by the model to verify whether it can accurately reflect the characteristics and laws of spatiotemporal data.

7. The method for extracting efficient features of power spatiotemporal data based on neural network according to claim 1 is characterized in that: The hybrid layer design includes hybrid layer architecture, training and optimization, and implementation and evaluation; The hybrid layer architecture includes: a feature extraction layer, a sequence processing layer, and a spiking neural network layer; The training and optimization include: hybrid loss function, back propagation and BPTT and hyperparameter adjustment; The implementation and evaluation include: Implement,hybrid layer architecture using deep learning frameworks; Evaluation,Experiments are conducted on standard datasets to evaluate the performance of the model and,compare it with models using only CNN, RNN, and SNN.

8. The method for efficient feature extraction of power spatiotemporal data based on neural network according to claim 7 is characterized in that: The feature extraction layer receives raw data, such as images and video frames, through the input layer, and uses multiple convolutional layers, pooling layers, and activation layers to extract spatial hierarchical features, which can then be used as input for subsequent layers; The sequence processing layer takes the feature sequence extracted by the CNN layer as input and uses the RNN layer to process the dependency in the time dimension; The spiking neural network layer takes the output of the RNN layer as input and further processes it through spiking neurons. The SNN layer simulates the discharge behavior of biological neurons to bring a more sensitive and efficient response to dynamic changes, and uses a threshold activation function to convert the output of the RNN layer into a spike sequence processed by the SNN layer.

9. The method for efficient feature extraction of power spatiotemporal data based on neural network according to claim 7 is characterized in that: The hybrid loss function is specifically: designing a loss function according to the task objective; The back propagation and BPTT are as follows: for the RNN layer and the CNN layer, the traditional back propagation algorithm and the back propagation over time algorithm are used for training, while the training of the SNN layer adopts the training algorithm based on the back propagation of the pulse time; The hyperparameter adjustment specifically includes: tuning the hyperparameters of the number of layers, number of neurons, and learning rate of the CNN layer, RNN layer, and SNN layer to maximize the model performance.

10. The method for efficient feature extraction of power spatiotemporal data based on neural network according to claim 8, characterized in that: The threshold activation function converts the input signal into an output signal, and its output depends on whether the input value exceeds a specific threshold. Its expression is as follows: Where θ is the threshold parameter set, x is the input value, and f(x) is the output value.

Images