Multi-monitoring-point grounding current prediction method and related device

Abstract

The invention belongs to a prediction method, provides a multi-monitoring-point grounding current prediction method and a related device, and aims to solve the technical problem that an existing multi-monitoring-point current prediction method is insufficient in prediction precision and robustness, and the multi-monitoring-point grounding current prediction method comprises the following steps: extracting a multi-dimensional reinforcement learning state from a sample after segmentation of a multi-monitoring-point continuous current time sequence; dynamically outputting an optimal time mask rate and an optimal space mask rate, enabling the optimal time mask rate and the optimal space mask rate to act on a sample to obtain an input tensor after space-time double mask, and extracting space-time representation from the input tensor after space-time double mask; according to the method, the prediction precision and robustness can be remarkably improved, and reliable technical support is provided for stable prediction and fault early warning of the multi-monitoring-point grounding current.

Description

Technical Field

[0001] This application pertains to a prediction method, specifically a method and related apparatus for predicting grounding current at multiple monitoring points. Background Technology

[0002] In fields such as power system operation and maintenance and smart grid dispatching, accurate prediction of current sequences from multiple monitoring points is the core foundation for ensuring stable system operation and enabling fault early warning. This type of current data exhibits significant spatiotemporal heterogeneity, with non-stationary fluctuations in the time dimension and complex spatial correlations. Effectively capturing these spatiotemporal characteristics and improving prediction accuracy and robustness are currently key technical challenges.

[0003] Existing multi-monitoring point current prediction technologies are mainly based on traditional time-series prediction models or single-dimensional deep learning models, which leads to the following drawbacks: First, fixed masking strategies cannot adapt to spatiotemporal heterogeneity, resulting in insufficient feature learning; second, single autoencoders coupled with spatiotemporal features result in low extraction accuracy; third, insufficient feature fusion leads to the loss of original detailed information; and fourth, the optimization objective is singular, resulting in weak generalization ability. Therefore, existing technologies are difficult to effectively address the spatiotemporal heterogeneity of current data and cannot meet the requirements of prediction accuracy and robustness in real-world scenarios. Summary of the Invention

[0004] This application addresses the technical problems of insufficient prediction accuracy and robustness of existing multi-monitoring point current prediction methods by providing a multi-monitoring point grounding current prediction method and related apparatus.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application proposes a method for predicting grounding current at multiple monitoring points, including: Acquire continuous current time series from multiple monitoring points, and divide the continuous current time series from multiple monitoring points into multiple segments, with each segment corresponding to a sample. Extract the multidimensional reinforcement learning state from each sample; Based on the state of multidimensional reinforcement learning, the optimal temporal mask rate and the optimal spatial mask rate are dynamically output; Apply the optimal temporal masking rate and the optimal spatial masking rate to the sample to obtain the input tensor after spatiotemporal double masking; Extract the spatiotemporal representation from the input tensor after the spatiotemporal double mask; The spatiotemporal representation is fused with the sample, and the fusion result is then used as input to obtain the prediction result through a downstream predictor.

[0006] Furthermore, the method for obtaining a continuous current time series from multiple monitoring points and dividing the continuous current time series from multiple monitoring points into multiple segments includes: Obtain continuous current time series from multiple monitoring points; Outlier removal and standardization were performed on the data in the continuous current time series of multiple monitoring points to obtain a standardized series. The standardized sequence is sampled using a sliding window method, with each window serving as a sample.

[0007] Furthermore, the states of the multidimensional reinforcement learning include temporal variance, temporal trend, spatial correlation, and data density.

[0008] Furthermore, the method for dynamically outputting the optimal temporal mask rate and the optimal spatial mask rate includes: A deep Q-network is used as the reinforcement learning agent, and the state of multidimensional reinforcement learning is used as the input of the deep Q-network. The action is selected through the ε-greedy policy, and the optimal temporal mask rate and the optimal spatial mask rate are dynamically output.

[0009] Furthermore, the method for applying the optimal temporal mask rate and the optimal spatial mask rate to the samples includes: Randomly mask the time steps of the sample according to the optimal time mask rate, and replace the masked positions with learnable time mask tokens; The monitoring points of the sample are randomly masked according to the spatial mask rate, and the masked positions are replaced with learnable spatial mask tokens.

[0010] Furthermore, the method for extracting spatiotemporal representations from the input tensor after the spatiotemporal double mask includes: Temporal and spatial representations are extracted from the input tensor after the spatiotemporal double mask using temporal autoencoders and spatial autoencoders, respectively. The temporal autoencoder and spatial autoencoder have the same structure, both including an encoder and a decoder. The encoder adopts a 4-layer Transformer with self-attention along the time dimension, and the decoder adopts a 1-layer Transformer and a regression layer. The temporal autoencoder and spatial autoencoder are trained using a joint reconstruction loss, which includes the temporal reconstruction loss of the temporal autoencoder and the spatial reconstruction loss of the spatial autoencoder.

[0011] Furthermore, the method for fusing the spatiotemporal representation by combining samples includes:

[0012] in, For the fusion result, As a representation of time, For spatial representation, The result is obtained by projecting the sample onto 96 dimensions using a single-layer MLP. The method for obtaining prediction results by using the fusion result as input through a downstream predictor includes: The downstream predictor uses a temporal Transformer structure to perform self-attention calculations on the fusion results to obtain preliminary prediction results; The initial prediction results are compressed into a dimension that matches the actual current data by using a fully connected layer to obtain a normalized current prediction value. By combining the pre-stored mean and standard deviation of each monitoring point, the normalized current prediction value is restored to the final prediction result in the original current unit.

[0013] Secondly, this application proposes a multi-monitoring-point grounding current prediction system, comprising: The segmentation module is used to acquire continuous current time series from multiple monitoring points and divide the continuous current time series from multiple monitoring points into multiple segments, with each segment forming a sample. The first extraction module is used to extract the multidimensional reinforcement learning state from each sample; The optimization module is used to dynamically output the optimal temporal mask rate and the optimal spatial mask rate based on the state of multidimensional reinforcement learning. The application module is used to apply the optimal temporal mask rate and the optimal spatial mask rate to the sample to obtain the input tensor after spatiotemporal double masking; The second extraction module is used to extract spatiotemporal representations from the input tensor after the spatiotemporal double mask; The prediction module is used to combine the spatiotemporal representation with the sample fusion, and then use the fusion result as input to obtain the prediction result through the downstream predictor.

[0014] Thirdly, this application proposes an electronic device, including: a memory and one or more processors; the memory is coupled to the processors; wherein the memory stores computer program code, the computer program code including computer instructions, and when the computer instructions are executed by the processor, the electronic device performs the steps of the above-described multi-monitoring point grounding current prediction method.

[0015] Fourthly, this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described multi-monitoring point grounding current prediction method.

[0016] Compared with the prior art, this application has the following beneficial effects: This application proposes a multi-monitoring point grounding current prediction method. It extracts multi-dimensional reinforcement learning states from segmented samples of continuous current time series from multiple monitoring points, dynamically outputting the optimal temporal and spatial masking rates. These optimal temporal and spatial masking rates are then applied to the samples to obtain a spatiotemporally double-masked input tensor. Spatiotemporal representations are extracted from this input tensor, fused, and the fused result is used as input to a downstream predictor to obtain the prediction result. This application introduces reinforcement learning to dynamically match the masking rate, replacing the traditional fixed masking strategy, which can adapt to the spatiotemporal heterogeneity of different samples. Simultaneously, through spatiotemporal representation extraction and fusion, the spatiotemporal correlation characteristics of the current data are fully explored, significantly improving prediction accuracy and robustness, providing reliable technical support for stable prediction and fault early warning of grounding current at multiple monitoring points.

[0017] This application also proposes a multi-monitoring point ground current prediction system, an electronic device, and a computer-readable storage medium, which possess all the advantages of the above-mentioned multi-monitoring point ground current prediction method. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a multi-monitoring point grounding current prediction method according to this application; Figure 2 This is a schematic diagram of an embodiment of the multi-monitoring point grounding current prediction method of this application; Figure 3 This is a comparison chart of prediction curves for the multi-monitoring-point grounding current prediction method of this application; Figure 4 This is a comparison of the prediction performance of the multi-monitoring point grounding current prediction method in this application under scenarios with similar short-term inputs but different future outputs; Figure 5 This is a comparison of the prediction performance of the multi-monitoring point grounding current prediction method in this application under scenarios with short-term input differences but future output convergence. Figure 6 This is a schematic diagram of a multi-monitoring point grounding current prediction system according to this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] In the description of the embodiments of this application, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this application. In addition, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] Furthermore, the use of the term "horizontal" does not imply that the component must be absolutely horizontal, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0025] In the description of the embodiments of this application, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0026] Existing multi-monitoring point current prediction technologies are mainly based on traditional time series prediction models or single-dimensional deep learning models, and can be specifically divided into the following two categories: The first category is prediction methods based on traditional time series analysis, such as the ARIMA model and exponential smoothing. These methods achieve short-term predictions by modeling the statistical characteristics of the current series over time. However, because they assume that the current series satisfies the premise of stationarity, they cannot adapt to the non-stationary fluctuation characteristics of current data in real-world scenarios. Furthermore, these methods do not consider the spatial correlation between different monitoring points, making it difficult to characterize the spatial heterogeneity of current data. This results in severely insufficient prediction accuracy in multi-monitoring scenarios, failing to meet the operational and maintenance needs of complex systems.

[0027] The second category comprises deep learning-based prediction methods, which have become a current research hotspot. These methods mainly include recurrent neural networks, Transformer models, and autoencoder-based representation learning models. Among them, autoencoder-based models improve the robustness of predictions by extracting and reconstructing features from spatiotemporal data, and are gradually being applied to the field of current prediction. However, existing autoencoder-based technical solutions still have the following significant drawbacks: (1) Fixed masking strategy with poor adaptability. Existing autoencoder models mostly use fixed temporal or spatial masking rates to process input data. However, in actual multi-monitoring current sequences, the spatiotemporal characteristics of different time segments and different combinations of monitoring points are significantly different. Fixed masking rates cannot adapt to this spatiotemporal heterogeneity, resulting in insufficient learning of local features by the model, which ultimately affects the quality of representation learning and prediction accuracy.

[0028] (2) Spatiotemporal feature coupling learning, resulting in low extraction accuracy: Existing models mostly use a single autoencoder to model both temporal and spatial features simultaneously, which leads to mutual interference between the fluctuation trend in the temporal dimension and the correlation in the spatial dimension, making it impossible to achieve accurate separation and efficient extraction of the two types of features. For example, when a short-term current change occurs at a certain monitoring point, it will mask the spatial collaborative features between other monitoring points, causing the representation learned by the model to have information confusion, making it difficult to support high-precision prediction.

[0029] (3) Insufficient feature fusion and loss of detailed information: In the feature learning process, existing solutions often rely too much on the abstract features extracted by the autoencoder and ignore the basic detailed information in the original input data. Although abstract features can depict the macroscopic laws of data, the lost detailed information may contain key representations of the equipment's operating status, leading to deviations between the model's prediction results and the actual current changes. This is especially true in scenarios that are sensitive to details, such as fault warnings, where warning delays or misjudgments are likely to occur.

[0030] (4) The model has a single optimization objective and weak generalization ability: Most existing models only use prediction loss as the optimization objective and have not established a collaborative optimization mechanism for representation learning quality and prediction accuracy. Although some schemes introduce reconstruction loss, they do not deeply associate it with the prediction task, which means that the features extracted by the autoencoder may not be suitable for downstream prediction needs. The model has insufficient generalization ability and its prediction performance drops significantly when faced with new monitoring point data or sudden load fluctuations.

[0031] Therefore, developing a multi-monitoring point current prediction technology that can adaptively adapt to spatiotemporal heterogeneity, accurately separate spatiotemporal features, fully integrate multi-source information, and achieve collaborative optimization has become an urgent technical problem to be solved in this field.

[0032] Based on the above, this application proposes a method and related device for predicting grounding current at multiple monitoring points. The following is a detailed description of this application in conjunction with embodiments and accompanying drawings.

[0033] like Figure 1 The diagram shown is a first flowchart of a multi-monitoring point grounding current prediction method according to this application, which may include: S101, acquire the continuous current time series of multiple monitoring points, and divide the continuous current time series of multiple monitoring points into multiple segments, with each segment corresponding to a sample.

[0034] Ground current prediction is essentially a time-series prediction task, requiring the transformation of continuous time-series data into fixed-length samples that can be processed by the model. By dividing the data into multiple segments to form samples, the temporal dependencies of the current data can be preserved, while providing basic units for subsequent sample-by-sample dynamic processing and feature extraction. In practical applications, the continuous current time series from multiple monitoring points can also be preprocessed to improve data quality and avoid interference from outliers and magnitude differences in subsequent model training.

[0035] S102 extracts the state of multidimensional reinforcement learning from each sample.

[0036] Reinforcement learning involves the iterative optimization of states, actions, and rewards, with states serving as the basis for agent decision-making. This application extracts multi-dimensional indicators that accurately characterize the spatiotemporal features of each sample, serving as the decision-making basis for the subsequent generation of dynamic mask rates by the reinforcement learning agent, achieving precise matching from sample features to the optimal mask rate. Therefore, it provides a comprehensive and accurate decision-making basis for the reinforcement learning agent, ensuring that the subsequently generated mask rate can adapt to the spatiotemporal heterogeneity of the samples. Furthermore, extracting multi-dimensional reinforcement learning states avoids decision biases caused by a single state dimension and enables personalized adaptation of processing strategies, laying the foundation for improving overall prediction accuracy.

[0037] S103, based on the state of multidimensional reinforcement learning, dynamically outputs the optimal temporal mask rate and the optimal spatial mask rate.

[0038] It should be noted that traditional prediction methods typically use a fixed masking rate to process all samples, which cannot adapt to the spatiotemporal heterogeneity of different samples. This application uses a reinforcement learning agent, taking multidimensional states as input and outputting temporal and spatial masking rates adapted to the current sample. Specifically, the temporal masking rate controls the proportion of time steps to be masked, and the spatial masking rate controls the proportion of monitoring points to be masked, achieving dynamic adaptation of the masking strategy. This application dynamically outputs the optimal temporal and spatial masking rates, replacing the fixed masking strategy, and achieving dynamic adaptation of the masking rate to the spatiotemporal features of the sample. This avoids the problem of losing key information due to too many masks or failing to uncover essential patterns due to too few masks, providing a more adaptable preprocessing strategy for subsequent accurate extraction of spatiotemporal features and improving the quality of feature extraction.

[0039] S104 applies the optimal temporal mask rate and the optimal spatial mask rate to the sample to obtain the input tensor after spatiotemporal double masking.

[0040] It should be noted that the purpose of masking is to force the model to reconstruct the global data from locally visible information, thereby uncovering the essential patterns in the data. This application uses masking operations in both temporal and spatial dimensions to partially mask the samples, generating a spatiotemporally masked input tensor. This provides suitable input data for subsequent autoencoder extraction of robust spatiotemporal features. This application enhances the model's ability to uncover spatiotemporal dependencies through dual-dimensional masking, avoiding overfitting of the model to the complete data.

[0041] S105, extract the spatiotemporal representation from the input tensor after the spatiotemporal double mask.

[0042] Spatiotemporal representation is an abstract expression of the spatiotemporal characteristics of current data, effectively capturing the essential patterns of the data. It enables decoupled learning of spatiotemporal features, improves the extraction accuracy of single-dimensional features, and avoids interference from spatiotemporal feature coupling.

[0043] S106, combine the spatiotemporal representation with the sample fusion, and then use the fusion result as input to obtain the prediction result through the downstream predictor.

[0044] Single temporal or spatial representations cannot fully cover all features of data; fusion enables the complementarity of multi-source information. The core role of the downstream predictor is to transform the fused high-dimensional representation into interpretable current prediction values, ultimately outputting raw unit prediction results that meet actual business needs. Representation fusion achieves complementarity between high-order abstract features and raw detailed features, avoiding information loss caused by single features and improving prediction accuracy.

[0045] like Figure 2 The diagram shown illustrates the principle of the multi-monitoring point grounding current prediction method of this application. Specifically, it may include: S201, Data Input and Preprocessing.

[0046] The original multi-monitoring point current sequences are converted into spatiotemporal tensors and reinforcement learning states, eliminating data noise and magnitude differences while preserving spatiotemporal heterogeneity features. The final output is a standardized sample set and the reinforcement learning state for each sample. and the average value of each monitoring point and standard deviation .

[0047] Specifically, this can be achieved through the following methods: (1) Input of raw data.

[0048] Input a continuous current time series from multiple monitoring points. The data it contains For the corresponding time Monitoring points current value .

[0049] (2) Outlier removal.

[0050] For the sequence at each monitoring point, if or ( For the first If the standard deviation of each monitoring point is considered, the data is determined to be an outlier, and the mean of the data from the preceding and following 5 steps is used to fill the gap.

[0051] (3) Standardization.

[0052] The mean was calculated independently for each monitoring point. and standard deviation Z-score standardization is performed using the following formula:

[0053] Z-score standardization is used to avoid the impact of differences in measurement ranges at different monitoring points on training.

[0054] (4) Sliding window sampling.

[0055] Set the input window length to 12, the prediction window length to 6, and the stride to 1. Extract samples from the standardized sequence. The input tensor for each sample is... The real label is .

[0056] It's important to note that the sliding window parameters can be configured comprehensively based on data characteristics, task requirements, and model capabilities. The goal is to ensure the window can fully capture key spatiotemporal features while adapting to the model's input and output dimensions. Specifically, the input window length ensures sufficient historical information to support prediction, while the prediction window length matches the prediction duration required by the business needs, thus guaranteeing prediction accuracy. The sliding step size is used to balance the number of samples and sample redundancy, adapting to different training or prediction scenarios.

[0057] (5) Reinforcement learning status extract.

[0058] Calculate the 4-dimensional reinforcement learning state for each sample That is, time variance Time trend Spatial correlation Data density.

[0059] Time variance This is the mean variance of all monitoring points over the current time window, i.e.:

[0060] Time trend The mean of the absolute values ​​of the first-order differences of all monitoring points within the current window is:

[0061] Spatial correlation This represents the mean of the Pearson correlation coefficients between all pairs of monitoring points.

[0062] Data density is the percentage of unfilled values ​​in the current sample.

[0063] S202, Dynamic dual mask rate for reinforcement learning agents.

[0064] Based on the reinforcement learning state S, the optimal temporal mask rate is dynamically output. and optimal spatial mask rate This replaces the fixed mask rate and adapts to the spatiotemporal features of different samples. That is, it takes the reinforcement learning state S as input and outputs a dynamic dual mask rate. .

[0065] Based on reinforcement learning state Dynamically output optimal time mask rate and spatial mask rate This method, instead of a fixed mask rate, can adapt to the spatiotemporal characteristics of different samples. Specifically, it can be achieved through the following methods: (1) Initialize DQN (Deep Q-Network) as a reinforcement learning agent. The DQN structure includes a 4-dimensional input layer → a 256-dimensional ReLU hidden layer → a 128-dimensional ReLU hidden layer → a 25-dimensional Q-value output layer. The action space is set to 25 combinations of dual mask rates. and All values ​​were selected from {0.1, 0.25, 0.4, 0.55, 0.7}.

[0066] The input layer corresponds to the reinforcement learning state of each sample, the hidden layer extracts features from the 4-dimensional state to explore the correlation between the reinforcement learning state and the optimal mask rate, and the output layer corresponds to a variety of preset dual mask rate combinations. The value of each dimension represents the estimated benefit that can be obtained by choosing this mask rate combination.

[0067] (2) Status input and action selection.

[0068] Strengthen learning state Input the DQN network and select actions using the ε-greedy policy: explore with probability ε and randomly select one action; utilize with probability 1-ε and select the action with the largest Q value output by the DQN.

[0069] For each sample, after inputting the 4-dimensional state into the DQN network, the agent selects an action. Exploration is performed with probability ε to try new mask rate combinations and prevent the agent from getting trapped in local optima. Utilization is performed with probability 1-ε, using learned experience to select the optimal solution; a higher estimated reward indicates a better prediction performance for the current state.

[0070] (3) Action output.

[0071] For the current sample, output a unique double mask rate. .

[0072] S203, a spatiotemporal double mask is applied to the input tensor.

[0073] By masking a portion of the time steps and monitoring points of the input tensor with a dynamic double mask rate, the autoencoder is forced to reconstruct the global data from locally visible information, thus enhancing its ability to capture spatiotemporal dependencies. The input consists of the sample input tensor and the dynamic double mask rate. The output is the input tensor after spatiotemporal double masking. It should be noted that the sample input tensor refers to the normalized sequence after the normalization process in step S201.

[0074] Specifically: (1) Execution time mask (T-Mask): based on time mask rate Randomly mask the time steps of the sample input tensor and replace the masked positions with learnable time mask tokens.

[0075] (2) Perform spatial mask (S-Mask): by spatial mask rate Randomly mask the monitoring points of the sample input tensor and replace the masked locations with learnable spatial mask tokens.

[0076] S204, dual autoencoder representation learning extracts robust spatiotemporal representations.

[0077] By learning temporal dependence and spatial correlation through temporal and spatial autoencoders respectively, a robust spatiotemporal representation that can characterize the current variation is extracted. The input is the input tensor after spatiotemporal double masking, and the output is the temporal representation. Spatial representation and joint reconstruction losses .

[0078] (1) Training of temporal autoencoder (T-MAE).

[0079] The time encoder employs a 4-layer Transformer with self-attention along the time dimension. The input tensor after being subjected to spatiotemporal double masking outputs a time representation with a dimension of 12×N×96.

[0080] The time decoder consists of a Transformer layer and a regression layer. It reconstructs the masked time step current value based on the time representation and the mask position.

[0081] Calculate the time reconstruction loss of T-MAE :

[0082] in, Represents the reconstruction value. The values ​​representing the masked monitoring points can be updated using gradient descent to update the T-MAE parameters.

[0083] It should be noted that the self-attention of the time autoencoder is calculated only along the time dimension, focusing on capturing the temporal variation characteristics of the current sequence at each monitoring point, avoiding interference from spatial dimension data, and enabling the output time representation to accurately characterize the temporal pattern of the current.

[0084] (2) Spatial autoencoder (S-MAE) training.

[0085] The encoder employs a 4-layer Transformer with self-attention only along the spatial dimension, and outputs a spatial representation with a dimension of 12×N×96 after inputting a mask tensor.

[0086] The spatial decoder consists of a Transformer layer and a regression layer. Based on the spatial representation and the mask location, it reconstructs the current value of the masked monitoring point.

[0087] Calculate the spatial reconstruction loss of S-MAE :

[0088] in, Represents the reconstruction value. The values ​​of the masked monitoring points are updated using gradient descent to update the S-MAE parameters.

[0089] It should be noted that the self-attention of the spatial autoencoder is calculated only along the spatial dimension, focusing on mining the current correlation characteristics of different monitoring points within the same time step, so that the output spatial representation can effectively reflect the spatial cooperative relationship between monitoring points.

[0090] (3) Calculate the joint reconstruction loss :

[0091] Based on the joint reconstruction loss, the parameters of the temporal autoencoder and spatial autoencoder are simultaneously fine-tuned.

[0092] Temporal and spatial representations have a natural synergistic relationship. Joint loss optimization forces the two autoencoders to adapt to each other during the learning process, avoiding the imbalance where the temporal encoder has a good fit but the spatial autoencoder has a large bias. This makes the final output dual representations complementary and consistent. The joint loss binds the optimization objectives of the two autoencoders, avoiding parameter redundancy and objective conflicts that occur during independent training. This accelerates the overall convergence speed of the model and reduces the risk of overfitting.

[0093] S205 represents the fusion and downstream prediction output results.

[0094] By integrating temporal and spatial representations with the original input information, a normalized predicted value is output from the downstream predictor and then denormalized to the actual current unit. Therefore, the predictor's input is the temporal representation. Spatial representation The input consists of the original tensor and the normalized true labels. The output consists of the normalized predicted values ​​and the prediction loss. And the final prediction results.

[0095] Specifically: (1) Representation fusion: The original input tensor is projected to 96 dimensions using a 1-layer MLP to obtain Computational Enhancement Representation :

[0096] (2) Downstream forecasting.

[0097] The downstream predictor can employ a temporal Transformer, with a 12-step augmented representation as input. The system outputs a 6-step prediction representation. Then, using a fully connected layer as the prediction head, the 6-step prediction representation is converted into a normalized current value with a dimension of 6×N×1.

[0098] Calculate the predicted loss :

[0099] in, For normalized predicted values, To normalize the true labels, the parameters of the predictor and dual autoencoder are updated using gradient descent. Based on the average value of each monitoring point and standard deviation The normalized predicted values ​​are converted into current values ​​in original units to obtain the final prediction result.

[0100] S206, Training and Testing.

[0101] The reinforcement learning agent can be optimized through model loss feedback, the model can be converged through batch training, and the generalization ability can be verified through testing. The input is the intermediate data related to each step, and the output is the training converged model and the final prediction result of new data.

[0102] Specific training methods may include: (1) Calculate the reward function The current state S and the selected action. The reward R and the next sample state S' are stored in the experience replay pool. 32 experience samples are sampled in a batch, and the target Q value is calculated. Minimize the Q-value error (MSE) and update the DQN parameters.

[0103] (2) Repeat steps S202-S206 for all sliding window samples to complete one round of training, iterate for 500 rounds, evaluate MAE with the validation set every 50 rounds, and save the model with the best performance on the validation set.

[0104] (3) For the new original current data of multiple monitoring points, repeat the cleaning, standardization, tensor construction and reinforcement learning state extraction of step S201. The trained reinforcement learning agent selects the optimal mask rate according to state S, and then obtains the ground current prediction result by inverse normalization after spatiotemporal masking, dual autoencoder extraction of representation, fusion representation and prediction.

[0105] This application provides a method for dynamically adapting to spatiotemporal heterogeneity and improving the prediction accuracy of ground current at multiple monitoring points. The original ground current time series from multiple monitoring points are cleaned, standardized, and sampled using a sliding window to generate a spatiotemporal tensor. A reinforcement learning state S containing temporal variance, temporal trend, spatial correlation, and data density is extracted. A reinforcement learning agent based on a DQN network dynamically outputs the temporal and spatial masking rates according to the state S. After applying a double mask to the spatiotemporal tensor, robust spatiotemporal representations are learned through a temporal autoencoder (T-MAE) and a spatial autoencoder (S-MAE). The fused representations are then output as normalized predicted values ​​by a downstream predictor and denormalized to actual current units. Simultaneously, reinforcement learning rewards are calculated based on model reconstruction loss and prediction loss to update the agent and optimize the masking rate. After multiple rounds of iterative training and convergence, the model is deployed to predict ground current in real time using new data.

[0106] Table 1

[0107] Table 1 presents a performance comparison of different models in the multi-monitoring point ground current prediction task, using the Mean Absolute Error (MAE) as the evaluation index. A smaller MAE value indicates higher prediction accuracy. The data in Table 1 shows that the traditional time series model (such as ARIMA) has an MAE of 0.412, indicating relatively limited prediction performance; the existing mainstream spatiotemporal prediction model (such as STAEformer) has an MAE of 0.201, demonstrating superior prediction accuracy; while the prediction model proposed in this application achieves an MAE of 0.178, significantly lower than other comparative models. These results demonstrate that, in the multi-monitoring point ground current prediction scenario, the prediction accuracy of the model in this application is superior to various comparative models in the existing technology, exhibiting outstanding technical performance.

[0108] like Figure 3 As shown, this is a comparison of the prediction curves of the multi-monitoring point grounding current prediction method of this application, demonstrating the fitting effect between the actual and predicted values. Figure 4 The figure shown is a comparison of the prediction performance of the multi-monitoring point grounding current prediction method in this application under scenarios with similar short-term inputs but different future outputs, demonstrating the model's ability to predict current changes over different time periods. Figure 5 The figure shown is a comparison of the prediction results of the multi-monitoring point grounding current prediction method of this application in a scenario with short-term input differences but future output convergence, which can verify the prediction stability of the model in the face of emergencies.

[0109] Existing technologies often employ fixed masking rates or single-dimensional masking strategies, which cannot address the dynamic changes (i.e., spatiotemporal heterogeneity) of current data from multiple monitoring points at different time steps and monitoring points. This application, however, utilizes a reinforcement learning agent to dynamically generate combinations of temporal and spatial masking rates based on a 4-dimensional state vector (temporal variance, temporal trend, spatial correlation, and data density). Each sample can obtain the optimal masking strategy, effectively adapting to the differences in data features under different scenarios and avoiding the problem of insufficient representation learning caused by fixed masks. Furthermore, existing spatiotemporal prediction methods often employ single autoencoders or hybrid masking mechanisms, making it difficult to separately focus on the dynamic trends in the temporal dimension and the correlation between monitoring points in the spatial dimension. This application employs a dual autoencoder architecture of T-MAE and S-MAE, respectively reconstructing time step information from tensors after temporal masking and reconstructing monitoring point information from tensors after spatial masking. This achieves decoupled modeling of spatiotemporal dependencies, resulting in more targeted and robust representations, overcoming the limitation of single encoders in simultaneously considering spatiotemporal features. Furthermore, existing masking methods often rely on manual experience or simple scheduling strategies to set the mask rate, lacking an adaptive optimization mechanism based on data features. This application constructs a reinforcement learning closed loop of "state awareness - action output - loss feedback - parameter update," using the weighted value of the model reconstruction loss and prediction loss as the reward signal to dynamically optimize the mask rate selection strategy. No manual parameter tuning is required throughout the process, significantly improving the model's adaptability to complex working conditions and training efficiency. Finally, existing multi-monitoring point grounding current prediction methods often rely on only a single feature or simple splicing representations, failing to fully utilize spatiotemporal abstract features and original data information. This application projects the original input to a dimension with the same origin as the spatiotemporal representation through MLP, and then performs feature fusion to form an enhanced representation. This retains the detailed information of the original data while incorporating robust spatiotemporal features refined by dual autoencoders, effectively strengthening the collaborative modeling capability among multiple monitoring points, and achieving prediction accuracy significantly better than traditional single-representation modeling methods.

[0110] like Figure 6 The diagram shown is a schematic of a multi-monitoring point ground current prediction system according to this application, which may include: The segmentation module is used to acquire continuous current time series from multiple monitoring points and divide the continuous current time series from multiple monitoring points into multiple segments, with each segment forming a sample. The first extraction module is used to extract the multidimensional reinforcement learning state from each sample; The optimization module is used to dynamically output the optimal temporal mask rate and the optimal spatial mask rate based on the state of multidimensional reinforcement learning. The application module is used to apply the optimal temporal mask rate and the optimal spatial mask rate to the sample to obtain the input tensor after spatiotemporal double masking; The second extraction module is used to extract spatiotemporal representations from the input tensor after the spatiotemporal double mask; The prediction module is used to combine the spatiotemporal representation with the sample fusion, and then use the fusion result as input to obtain the prediction result through the downstream predictor.

[0111] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of each block is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple blocks may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.

[0112] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.

[0113] This application also provides an electronic device, which may include one or more processors, memory and communication interfaces.

[0114] The memory, communication interface, and processor are coupled together. For example, the memory, communication interface, and processor can be coupled together via a bus.

[0115] The communication interface is used for data transmission with other devices. The memory stores computer program code. This computer program code includes computer instructions, which, when executed by the processor, cause the electronic device to perform the steps of the multi-monitoring-point grounding current prediction method described above.

[0116] The processor can be a processor or controller, such as a Central Processing Unit (CPU), a general-purpose processor, a Digital Signal Processor (DSP), an Application-Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with this disclosure. The processor can also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The processor can be used to support an electronic device in performing the method steps provided in the above embodiments.

[0117] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. These buses can be categorized as address buses, data buses, control buses, etc.

[0118] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the above-described multi-monitoring point grounding current prediction method.

[0119] The computer-readable storage media involved in this application include random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage media known in the art.

[0120] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A multi-monitoring-point ground current prediction method characterized by, The method comprises the following steps: obtaining a multi-monitoring-point continuous current time sequence, and dividing the multi-monitoring-point continuous current time sequence into multiple segments, each segment corresponding to form a sample; extracting a multi-dimensional reinforcement learning state from each sample respectively; based on the multi-dimensional reinforcement learning state, dynamically outputting an optimal time mask rate and an optimal space mask rate; making the optimal time mask rate and the optimal space mask rate act on the sample to obtain an input tensor after time-space double masking; extracting a time-space representation from the input tensor after time-space double masking; combining the sample and fusing the time-space representation, and taking the fusion result as an input to obtain a prediction result through a downstream predictor.

2. The method of claim 1, wherein, The method for obtaining a multi-monitoring-point continuous current time sequence and dividing the multi-monitoring-point continuous current time sequence into multiple segments comprises the following steps: obtaining a multi-monitoring-point continuous current time sequence; performing outlier rejection and standardization processing on the data in the multi-monitoring-point continuous current time sequence to obtain a standardized sequence; performing sliding window sampling on the standardized sequence, and taking each window as a sample.

3. The method of claim 1, wherein, The multi-dimensional reinforcement learning state comprises a time variance, a time trend, a space correlation and a data density.

4. The method of claim 1, wherein, The method for dynamically outputting an optimal time mask rate and an optimal space mask rate comprises the following steps: adopting a deep Q network as a reinforcement learning agent, taking the multi-dimensional reinforcement learning state as an input of the deep Q network, selecting an action through an ε-greedy strategy, and dynamically outputting an optimal time mask rate and an optimal space mask rate.

5. The method of claim 1, wherein, The method for making the optimal time mask rate and the optimal space mask rate act on the sample comprises the following steps: randomly masking a time step of the sample according to the optimal time mask rate, and replacing the masked position with a learnable time mask token; randomly masking a monitoring point of the sample according to the space mask rate, and replacing the masked position with a learnable space mask token.

6. The method of claim 1, wherein, The method for extracting a time-space representation from the input tensor after time-space double masking comprises the following steps: respectively adopting a time autoencoder and a space autoencoder to extract a time representation and a space representation from the input tensor after time-space double masking; the time autoencoder and the space autoencoder have the same structure, and each comprises an encoder and a decoder; the encoder adopts a 4-layer Transformer with time dimension self-attention, and the decoder adopts a 1-layer Transformer and a regression layer; the time autoencoder and the space autoencoder adopt a joint reconstruction loss during training, and the joint reconstruction loss comprises a time reconstruction loss of the time autoencoder and a space reconstruction loss of the space autoencoder.

7. The method of claim 1, wherein, The method for combining the sample and fusing the time-space representation comprises the following steps: wherein, is the fusion result, is the temporal representation, is the spatial representation, is the result of projecting the samples to 96 dimensions by a 1-layer MLP; The method for taking the fusion result as an input to obtain a prediction result through a downstream predictor comprises the following steps: the downstream predictor adopts a time series Transformer structure to perform self-attention calculation on the fusion result to obtain a preliminary prediction result; a 1-layer fully connected layer is used to compress the preliminary prediction result into a dimension matching the real current data to obtain a normalized current prediction value; in combination with pre-stored mean values and standard deviations of each monitoring point, the normalized current prediction value is restored to an ultimate prediction result in an original current unit.

8. A multi-monitoring-point ground current prediction system, comprising: The method comprises the following steps: The segmentation module is configured to obtain a multi-monitoring-point continuous current time sequence, and divide the multi-monitoring-point continuous current time sequence into multiple segments, each of which corresponds to a sample; The first extraction module is configured to extract a multi-dimensional reinforcement learning state from each sample respectively; The optimization module is configured to dynamically output an optimal time mask rate and an optimal space mask rate based on the multi-dimensional reinforcement learning state; The action module is configured to make the optimal time mask rate and the optimal space mask rate act on the sample to obtain a time-space double-masked input tensor; The second extraction module is configured to extract a time-space representation from the time-space double-masked input tensor; The prediction module is configured to combine the sample with the time-space representation, and take the fusion result as an input to obtain a prediction result through a downstream predictor.

9. An electronic device, comprising: The electronic device comprises: a memory and one or more processors; the memory is coupled with the processor; wherein the memory stores computer program code, the computer program code comprises computer instructions, when the computer instructions are executed by the processor, the electronic device executes the steps of the multi-monitoring-point grounding current prediction method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the multi-monitoring-point grounding current prediction method according to any one of claims 1-7.

Images