Power dispatching method and device fusing multi-modal data and dual verification, electronic equipment and storage medium

Abstract

The invention discloses an electric power dispatching method and device fusing multi-modal data and dual verification, electronic equipment and a storage medium, and belongs to the field of intelligent dispatching of an electric power system.The method comprises the steps that multi-modal dynamic data, including equipment images, load time sequences, logs, dispatching instructions, meteorological early warning and electricity price data, of an electric power system are obtained; visual, time sequence and semantic features are extracted, cross-modal alignment is carried out, and a joint representation vector is generated; based on the vector and the real-time data, an initial scheduling strategy is generated through a reinforcement learning model; and through historical scheme matching and physical constraint verification, determining a target scheduling strategy and executing scheduling. Therefore, by implementing the method and the device, the problems of incomplete power grid state perception caused by insufficient multi-modal data fusion and low reliability caused by lack of automatic credibility verification of the AI scheduling strategy in the prior art can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent dispatching of power systems, and in particular to a power dispatching method and device fusing multi-modal data and double verification, an electronic device and a storage medium. BACKGROUND

[0002] Intelligent dispatching of power systems is the core of ensuring safe and economic operation of power grids. With the development of artificial intelligence technology, using multi-modal data (such as device images, load curves, operation and maintenance logs, etc.) for comprehensive decision-making has become a key direction for improving dispatching accuracy. However, existing intelligent dispatching technologies still face two major bottlenecks in practice: Firstly, there is insufficient fusion of multi-modal data. Many technical solutions still use data in isolation, for example, only analyzing SCADA time series data, but failing to effectively fuse abnormal hot spots in device infrared images or fault texts in operation and maintenance logs. This "fragmentation" of data prevents AI models from forming a comprehensive and unified understanding of the current real state of the power grid, which poses a hidden danger for subsequent decision-making. Secondly, there is a lack of verification of the credibility of AI decision-making. On the basis of incomplete data perception, the dispatching strategies generated by AI models (such as reinforcement learning) often lack automated verification mechanisms. These strategies may deviate from historical optimal operation examples, or even may violate critical physical operation constraints of the power grid. This coexistence of "incomplete perception" and "black box decision-making" results in low reliability of AI-generated strategies, which requires extensive manual review and seriously hinders the closed-loop application of intelligent dispatching. SUMMARY

[0003] Embodiments of the present application provide a power dispatching method and device fusing multi-modal data and double verification, an electronic device and a storage medium, which can solve the problem of incomplete perception of the state of the power grid due to insufficient fusion of multi-modal data in the prior art, and the low reliability of AI dispatching strategies due to the lack of automated credibility verification.

[0004] An embodiment of the present application provides a power dispatching method fusing multi-modal data and double verification, comprising: obtaining multi-modal dynamic data of a power system, the multi-modal dynamic data comprising device image data, power grid load time series data, operation and maintenance log data, dispatching instruction data, weather warning data and real-time electricity price data; extracting visual features from the device image data to capture local abnormal regions and generating visual features; encoding time series features from the power grid load time series data and generating time series features; analyzing semantic features from the operation and maintenance log data and the dispatching instruction data and generating semantic features; and performing cross-modal alignment processing on the visual features, the time series features and the semantic features to generate a unified joint representation vector; According to the unified joint representation vector, real-time electricity price data and meteorological warning data, an initial scheduling strategy is generated through a preset reinforcement learning model; The initial scheduling strategy is matched with a preset historical optimal scheme database to obtain a similarity matching result, and the initial scheduling strategy is checked for compliance with a preset physical constraint rule database to obtain a compliance checking result; Based on the similarity matching result and the compliance checking result, a target scheduling strategy is determined; According to the target scheduling strategy, the power system is scheduled.

[0005] Further, the device image data is subjected to visual feature extraction to capture local abnormal regions, and visual features are generated, including: The device image data is input into a preset visual feature extraction model, so that the visual feature extraction model generates visual features according to the device image data; The visual feature extraction model is generated by fine-tuning a pre-trained Vision Transformer model as follows: An electric power system visual image training data set is obtained; wherein the electric power system visual image training data set includes a plurality of image-text pairs, each image-text pair including an electric power system device image and a description text describing the running state of the electric power system device; The electric power system visual image training data set is divided into a plurality of batches of first training samples according to a preset batch size; Each batch of first training samples is input into the pre-trained Vision Transformer model for iterative fine-tuning until a preset number of training rounds is reached; The pre-trained Vision Transformer model generates corresponding image embedding vectors and text embedding vectors based on the electric power system device images and description texts in the current batch when receiving each batch of first training samples; Based on a contrastive learning loss function, the similarity between the matched image embedding vectors and text embedding vectors is maximized to calculate a first loss function value; Using a preset optimizer, the learnable parameters in the pre-trained Vision Transformer model are updated according to the first loss function value, the training of the pre-trained Vision Transformer model is completed, and the trained pre-trained Vision Transformer model is denoted as the visual feature extraction model.

[0006] Furthermore, the step of encoding the time-series data of the power grid load to generate time-series features includes: The power grid load time series data is input into a preset time series feature encoding model so that the time series feature encoding model generates time series features based on the power grid load time series data; The temporal feature encoding model is generated through the following training: Obtain a power grid time-series training dataset; wherein, the power grid time-series training dataset includes several time-series training samples, each of the time-series training samples includes power grid load data for a historical period, renewable energy output curve data, and corresponding future period real load fluctuation labels; The power grid time series training dataset is divided into several batches of second training samples according to a preset batch size; The second training samples of each batch are sequentially input into the temporal feature encoding model for iterative training until the preset number of training rounds is reached; The time-series feature coding model outputs predicted load fluctuations based on the historical grid load data and new energy output curve data of the current batch when it receives a batch of second training samples. Based on the mean squared error loss function, the second loss function value is calculated according to the predicted load fluctuation and the actual load fluctuation label corresponding to the time series training sample; Using a preset optimizer, the learnable parameters in the temporal feature coding model are updated according to the second loss function value, thereby completing the training of the temporal feature coding model.

[0007] Furthermore, the semantic feature parsing of the operation and maintenance log data and scheduling instruction data to generate semantic features includes: The operation and maintenance log data and scheduling instruction data are input into a preset semantic feature parsing model, so that the semantic feature parsing model generates semantic features based on the operation and maintenance log data and scheduling instruction data; The semantic feature parsing model is obtained by fine-tuning a pre-trained BERT model as follows: Obtain a text training dataset in the power industry; wherein, the text training dataset in the power industry includes several text samples in the power industry, and the text samples in the power industry include historical operation and maintenance log data and scheduling instruction data; The power industry text training dataset is divided into several batches of third training samples according to a preset batch size; The third training samples from each batch are sequentially input into the general pre-trained BERT model for iterative fine-tuning until the preset number of training rounds is reached. The general pre-trained BERT model performs a masking language model task on the text samples in the current batch when it receives the third training sample in each batch, and outputs the prediction results of the masked words. Based on a cross-entropy loss function, the value of a third loss function is calculated according to the prediction result of the masked word and the corresponding real word in the text sample. Using a preset optimizer, the learnable parameters in the general pre-trained BERT model are updated according to the value of the third loss function to complete the fine-tuning training of the BERT model, and the fine-tuned BERT model is recorded as the semantic feature parsing model.

[0008] Furthermore, the step of performing cross-modal alignment processing on the visual features, the temporal features, and the semantic features to generate a unified joint representation vector includes: The visual features, temporal features, and semantic features are input into a preset cross-modal attention layer, so that the cross-modal attention layer fuses the visual features, temporal features, and semantic features to generate a unified joint representation vector; The cross-modal attention layer is obtained by training a joint model through the following steps: Obtain a power dispatch scenario dataset; wherein, the power dispatch scenario dataset includes several power dispatch scenario training samples, each power dispatch scenario training sample includes a set of equipment image training data, power grid load time series training data, operation and maintenance log training data and dispatch instruction training data acquired at the same time; the joint model includes the visual feature extraction model, the time series feature encoding model, the semantic feature parsing model and a cross-modal attention layer to be trained; The power dispatch scenario dataset is divided into several batches of fourth training samples according to a preset batch size; The fourth training samples from each batch are sequentially input into the joint model for iterative training until the preset number of training rounds is reached. When the joint model receives the fourth training sample in each batch, it calls the visual feature extraction model, the temporal feature encoding model and the semantic feature parsing model respectively, and generates training visual features, training temporal features and training semantic features based on the equipment image training data, power grid load temporal training data, operation and maintenance log training data and scheduling instruction training data in the current batch of the fourth training sample. The generated visual features, temporal features, and semantic features for training are input into the cross-modal attention layer to generate a unified joint representation vector for training. Based on the cross-modal alignment loss function, the fourth loss function value is calculated according to the unified joint representation vector used for training; wherein, the cross-modal alignment loss function is used to bring the distance between each interrelated feature in the fourth training sample in the unified joint representation vector space closer, and to push away the distance between unrelated features. Using a preset optimizer, the learnable parameters in the cross-modal attention layer are updated according to the fourth loss function value, thereby completing the training of the cross-modal attention layer.

[0009] Furthermore, the initial scheduling strategy is generated based on the unified joint representation vector, real-time electricity price data, and meteorological early warning data through a preset reinforcement learning model, including: The unified joint representation vector, the real-time electricity price data, and the meteorological early warning data are combined to construct a current state vector; The current state vector is input into the reinforcement learning model so that the reinforcement learning model outputs an action containing the energy storage charging and discharging power setting value and the thermal power unit output adjustment command based on the current state vector, as the initial scheduling strategy; The reinforcement learning model was obtained through offline training in a power system simulation environment as follows: The reinforcement learning model is iteratively trained in the power system simulation environment until a preset number of training rounds are reached, thus completing the training of the reinforcement learning model. In each iteration, a simulation state vector from the power system simulation environment is obtained as the current simulation state. The current simulation state is input into the reinforcement learning model so that the reinforcement learning model generates a simulation action based on the current simulation state; The simulation action is submitted to the power system simulation environment, so that the power system simulation environment performs simulation execution according to the simulation action, and generates a simulation state at the next moment and a simulation execution result. Based on the simulation results, and using a multi-objective reward function, a reward value is calculated. An experience tuple is constructed based on the current simulation state, the simulation action, the reward value, and the simulation state at the next moment. A proximal policy optimization algorithm is used to update the learnable parameters in the reinforcement learning model based on the empirical tuple, thereby completing the update of the reinforcement learning model in this iteration.

[0010] Furthermore, the power system simulation environment is constructed in the following manner: Obtain a power system physical simulator; Acquire topology data of the power system, operating parameter data of power system equipment, and physical operating constraints of the power system; The topology data, the operating parameter data, and the physical operating constraints are configured into the power system physical simulator to construct the power system simulation environment.

[0011] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0012] One embodiment of the present invention provides a power dispatching device that integrates multimodal data and dual verification, including: a power multimodal dynamic data acquisition module, a unified joint representation vector generation module, an initial dispatching strategy generation module, a target dispatching strategy determination module, and a power system dispatching module; The power multimodal dynamic data acquisition module is used to acquire multimodal dynamic data of the power system. The multimodal dynamic data includes equipment image data, power grid load time series data, operation and maintenance log data, dispatch instruction data, meteorological early warning data, and real-time electricity price data. The unified joint representation vector generation module is used to extract visual features from equipment image data to capture local abnormal areas and generate visual features; encode time-series load data to generate time-series features; parse semantic features from operation and maintenance log data and scheduling instruction data to generate semantic features; and perform cross-modal alignment processing on the visual features, the time-series features and the semantic features to generate a unified joint representation vector. The initial scheduling strategy generation module is used to generate an initial scheduling strategy based on a unified joint representation vector, real-time electricity price data, and meteorological early warning data, using a preset reinforcement learning model. The target scheduling strategy determination module is used to perform similarity matching calculation between the initial scheduling strategy and a preset historical optimal solution database to obtain a similarity matching result; and to perform compliance verification between the initial scheduling strategy and a preset physical constraint rule database to obtain a compliance verification result. The power system dispatching module is used to dispatch the power system according to the target dispatching strategy.

[0013] Based on the above method embodiments, the present invention provides corresponding electronic device embodiments.

[0014] One embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the power dispatching method integrating multimodal data and dual verification as described in any of the above-described method embodiments.

[0015] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments.

[0016] One embodiment of the present invention provides a storage medium storing a computer program thereon, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the power dispatching method integrating multimodal data and dual verification as described in any of the above-described method embodiments.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a power dispatching method, apparatus, electronic device, and storage medium that integrates multimodal data and dual verification. The method acquires multimodal dynamic data of the power system, including equipment image data, grid load time-series data, operation and maintenance log data, dispatching instruction data, meteorological early warning data, and real-time electricity price data. Visual features are extracted from the equipment image data to identify local anomaly areas, generating visual features. Time-series features are encoded from the grid load time-series data, generating time-series features. Semantic parsing is performed on the operation and maintenance log data and dispatching instruction data to generate semantic features. Visual, time-series, and semantic features are aligned across modalities to generate a unified joint representation vector. Based on this representation vector, real-time electricity price data, and meteorological early warning data, an initial dispatching strategy is generated using a preset reinforcement learning model. The initial dispatching strategy is matched for similarity with a historical optimal solution database and verified for compliance with a physical constraint rule database. A target dispatching strategy is determined based on the matching and verification results, and power system dispatching is executed accordingly.

[0018] This invention first extracts features and performs cross-modal alignment processing on equipment image data, power grid load time-series data, and operation and maintenance log data to generate a unified joint representation vector, thereby achieving a comprehensive and accurate perception of the power grid status. Based on this, the invention further performs dual verification on the generated initial strategy: compliance verification with a pre-set physical constraint rule database ensures the strategy's safety baseline; and similarity matching with a pre-set historical best solution database ensures the strategy's optimality. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a power dispatching method that integrates multimodal data and dual verification, provided by an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of a power dispatching device that integrates multimodal data and dual verification according to an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, to address the problems in existing technologies where insufficient multimodal data fusion leads to incomplete power grid state perception and the lack of automated reliability verification in AI scheduling strategies results in low reliability, an embodiment of the present invention provides a power scheduling method that integrates multimodal data and dual verification, comprising at least the following steps: Step S1: Obtain multimodal dynamic data of the power system. The multimodal dynamic data includes equipment image data, grid load time series data, operation and maintenance log data, dispatch instruction data, meteorological early warning data, and real-time electricity price data. Specifically, the system first needs to access and aggregate multi-source heterogeneous dynamic data in real time from various subsystems of the power system. The equipment image data is collected through cameras and infrared sensors deployed on substations, line towers, or key equipment (such as transformers and circuit breakers), mainly including visible light images reflecting the equipment's appearance and infrared thermal images reflecting the equipment's operating temperature distribution. The grid load time-series data comes from power dispatch automation systems (such as SCADA systems), providing real-time load curves for the entire network and key sections, historical load data, and actual power output data from renewable energy power plants.

[0023] Both the operation and maintenance log data and the dispatch instruction data are text-based. The operation and maintenance log data records historical equipment fault descriptions, inspection information, defect reports, and maintenance status; the dispatch instruction data contains operational instructions issued by dispatchers during operation. These instructions can be structured text records or unstructured voice records requiring speech recognition conversion. Furthermore, the meteorological early warning data is used to obtain early warning information from the meteorological system regarding extreme weather (such as strong winds and thunderstorms) that may affect line safety or renewable energy output. The real-time electricity price data reflects the current supply and demand relationship and price signals in the electricity market. By comprehensively acquiring these six heterogeneous data types, a data foundation is laid for subsequently building accurate and unified power grid status awareness.

[0024] Step S2: Extract visual features from equipment image data to capture local abnormal areas and generate visual features; encode time-series data of power grid load to generate time-series features; parse semantic features from operation and maintenance log data and scheduling instruction data to generate semantic features; perform cross-modal alignment processing on the visual features, the time-series features and the semantic features to generate a unified joint representation vector. In a preferred embodiment, the step of extracting visual features from device image data to capture local abnormal regions and generating visual features includes: The device image data is input into a preset visual feature extraction model so that the visual feature extraction model generates visual features based on the device image data; The visual feature extraction model is generated by fine-tuning a pre-trained Vision Transformer model through the following training: Obtain a power system visual image training dataset; wherein, the power system visual image training dataset includes several image-text pairs, each image-text pair including an image of power system equipment and a descriptive text used to describe the operating status of the power system equipment; The power system visual image training dataset is divided into several batches of first training samples according to a preset batch size; The first training samples of each batch are sequentially input into the pre-trained Vision Transformer model for iterative fine-tuning until the preset number of training rounds is reached; The pre-trained Vision Transformer model generates corresponding image embedding vectors and text embedding vectors based on the power system equipment images and descriptive text in the current batch when it receives the first training samples of each batch. Based on a contrastive learning loss function, the first loss function value is calculated with the goal of maximizing the similarity between the matched image embedding vector and the text embedding vector. Using a preset optimizer, the learnable parameters in the pre-trained Vision Transformer model are updated according to the first loss function value to complete the training of the pre-trained Vision Transformer model, and the trained pre-trained Vision Transformer model is recorded as the visual feature extraction model.

[0025] In a preferred embodiment, the step of encoding the time-series data of the power grid load to generate time-series features includes: The power grid load time series data is input into a preset time series feature encoding model so that the time series feature encoding model generates time series features based on the power grid load time series data; The temporal feature encoding model is generated through the following training: Obtain a power grid time-series training dataset; wherein, the power grid time-series training dataset includes several time-series training samples, each of the time-series training samples includes power grid load data for a historical period, renewable energy output curve data, and corresponding future period real load fluctuation labels; The power grid time series training dataset is divided into several batches of second training samples according to a preset batch size; The second training samples of each batch are sequentially input into the temporal feature encoding model for iterative training until the preset number of training rounds is reached; The time-series feature coding model outputs predicted load fluctuations based on the historical grid load data and new energy output curve data of the current batch when it receives a batch of second training samples. Based on the mean squared error loss function, the second loss function value is calculated according to the predicted load fluctuation and the actual load fluctuation label corresponding to the time series training sample; Using a preset optimizer, the learnable parameters in the temporal feature coding model are updated according to the second loss function value, thereby completing the training of the temporal feature coding model.

[0026] In a preferred embodiment, the step of performing semantic feature parsing on the operation and maintenance log data and scheduling instruction data to generate semantic features includes: The operation and maintenance log data and scheduling instruction data are input into a preset semantic feature parsing model, so that the semantic feature parsing model generates semantic features based on the operation and maintenance log data and scheduling instruction data; The semantic feature parsing model is obtained by fine-tuning a pre-trained BERT model as follows: Obtain a text training dataset in the power industry; wherein, the text training dataset in the power industry includes several text samples in the power industry, and the text samples in the power industry include historical operation and maintenance log data and scheduling instruction data; The power industry text training dataset is divided into several batches of third training samples according to a preset batch size; The third training samples from each batch are sequentially input into the general pre-trained BERT model for iterative fine-tuning until the preset number of training rounds is reached. The general pre-trained BERT model performs a masking language model task on the text samples in the current batch when it receives the third training sample in each batch, and outputs the prediction results of the masked words. Based on a cross-entropy loss function, the value of a third loss function is calculated according to the prediction result of the masked word and the corresponding real word in the text sample. Using a preset optimizer, the learnable parameters in the general pre-trained BERT model are updated according to the value of the third loss function to complete the fine-tuning training of the BERT model, and the fine-tuned BERT model is recorded as the semantic feature parsing model.

[0027] In a preferred embodiment, the step of performing cross-modal alignment processing on the visual features, the temporal features, and the semantic features to generate a unified joint representation vector includes: The visual features, temporal features, and semantic features are input into a preset cross-modal attention layer, so that the cross-modal attention layer fuses the visual features, temporal features, and semantic features to generate a unified joint representation vector; The cross-modal attention layer is obtained by training a joint model through the following steps: Obtain a power dispatch scenario dataset; wherein, the power dispatch scenario dataset includes several power dispatch scenario training samples, each power dispatch scenario training sample includes a set of equipment image training data, power grid load time series training data, operation and maintenance log training data and dispatch instruction training data acquired at the same time; the joint model includes the visual feature extraction model, the time series feature encoding model, the semantic feature parsing model and a cross-modal attention layer to be trained; The power dispatch scenario dataset is divided into several batches of fourth training samples according to a preset batch size; The fourth training samples from each batch are sequentially input into the joint model for iterative training until the preset number of training rounds is reached. When the joint model receives the fourth training sample in each batch, it calls the visual feature extraction model, the temporal feature encoding model and the semantic feature parsing model respectively, and generates training visual features, training temporal features and training semantic features based on the equipment image training data, power grid load temporal training data, operation and maintenance log training data and scheduling instruction training data in the current batch of the fourth training sample. The generated visual features, temporal features, and semantic features for training are input into the cross-modal attention layer to generate a unified joint representation vector for training. Based on the cross-modal alignment loss function, the fourth loss function value is calculated according to the unified joint representation vector used for training; wherein, the cross-modal alignment loss function is used to bring the distance between each interrelated feature in the fourth training sample in the unified joint representation vector space closer, and to push away the distance between unrelated features. Using a preset optimizer, the learnable parameters in the cross-modal attention layer are updated according to the fourth loss function value, thereby completing the training of the cross-modal attention layer.

[0028] Specifically, the core objective of this step is to transform the heterogeneous and multi-source data obtained in S1 into a unified and computable feature vector through their respective professional feature extraction models, and finally fuse them into a unified joint representation vector that can comprehensively represent the current power grid state.

[0029] In one embodiment, the step of extracting visual features from device image data to capture local abnormal regions and generating visual features includes: The device image data is input into a preset visual feature extraction model, which generates visual features based on the image data. Specifically, the visual feature extraction model can employ a VisionTransformer (ViT) architecture. This model excels at segmenting and encoding device images (including visible light images or infrared thermal images), effectively capturing spatial features and local anomalies in the image, such as identifying minute cracks in insulators or localized overheating distributions in transformers. The visual feature extraction model is generated by fine-tuning a pre-trained VisionTransformer model through the following steps: Obtain a power system visual image training dataset; wherein, the power system visual image training dataset includes several image-text pairs, each image-text pair includes an image of a power system device and a descriptive text used to describe the operating status of the power system device; a specific image-text pair may be "infrared image of device A" and its corresponding descriptive text "the temperature of the connection point of device A is too high".

[0030] The power system visual image training dataset is divided into several batches of first training samples according to a preset batch size; the first training samples of each batch are sequentially input into the pre-trained Vision Transformer model for iterative fine-tuning until a preset number of training rounds is reached; wherein, when the pre-trained Vision Transformer model receives a batch of first training samples, it generates corresponding image embedding vectors and text embedding vectors based on the power system equipment images and descriptive text in the current batch; based on a contrastive learning loss function, the first loss function value is calculated with the goal of maximizing the similarity between the matched image embedding vector and the text embedding vector; The goal of this contrastive learning loss function (such as CLIP-style loss) is to narrow the distance between matching image-text pairs in the feature space while widening the distance between mismatched image-text pairs. Using a pre-defined optimizer, the learnable parameters in the pre-trained Vision Transformer model are updated based on the first loss function value, completing the training of the pre-trained Vision Transformer model. The trained pre-trained Vision Transformer model is then designated as the visual feature extraction model.

[0031] In one embodiment, the step of encoding the time-series data of the power grid load to generate time-series features includes: Before inputting time series data into the model, preprocessing can be performed. Specifically, the Time Series Decomposition (STL) method (Seasonal and Trend decomposition using Loess) is used to decompose the original load curve data into trend, periodic, and residual terms, thereby eliminating the interference of periodic noise and standardizing the data.

[0032] The power grid load time series data (referring to the preprocessed data) is input into a preset time series feature coding model so that the time series feature coding model generates time series features based on the power grid load time series data; The time-series feature encoding model can employ the Informer model, whose ProbSparse self-attention mechanism can efficiently process long-series time-series data, effectively predict future (e.g., the next 24 hours) load fluctuations and new energy output curves, and generate corresponding feature embedding vectors. The time-series feature encoding model is generated through the following training: A power grid time-series training dataset is obtained; wherein the power grid time-series training dataset includes several time-series training samples, each of which includes power grid load data for a historical period, renewable energy output curve data, and a corresponding future period's actual load fluctuation label; the power grid time-series training dataset is divided into several batches of second training samples according to a preset batch size; each batch of second training samples is sequentially input into the time-series feature encoding model for iterative training until a preset number of training rounds is reached; wherein, when the time-series feature encoding model receives a batch of second training samples, it outputs a predicted load fluctuation based on the power grid load data and renewable energy output curve data for the historical period in the current batch; based on the mean squared error (MSE) loss function, a second loss function value is calculated according to the predicted load fluctuation and the actual load fluctuation label corresponding to the time-series training sample; using a preset optimizer, the learnable parameters in the time-series feature encoding model are updated according to the second loss function value to complete the training of the time-series feature encoding model.

[0033] In one embodiment, the step of performing semantic feature parsing on operation and maintenance log data and scheduling instruction data to generate semantic features includes: The operation and maintenance log data and scheduling instruction data are input into a preset semantic feature parsing model, which generates semantic features based on the operation and maintenance log data and scheduling instruction data. The semantic feature parsing model can be a BERT model. By fine-tuning on power industry text, the model can accurately understand and extract key semantic information from power industry-specific terms, equipment IDs, fault types, and operation instructions. The semantic feature parsing model is obtained by fine-tuning a pre-trained BERT model through the following steps: acquiring a power industry text training dataset; wherein the power industry text training dataset includes several power industry text samples, including historical operation and maintenance log data and scheduling instruction data; dividing the power industry text training dataset into several batches of third training samples according to a preset batch size; sequentially inputting each batch of third training samples into the general pre-trained BERT model for iterative fine-tuning until a preset number of training rounds is reached; wherein, when the general pre-trained BERT model receives a batch of third training samples, it executes a masked language model on the text samples in the current batch. The task is to output the prediction result of the masked word; based on a cross-entropy loss function, the third loss function value is calculated according to the prediction result of the masked word and the corresponding real word in the text sample; using a preset optimizer, the learnable parameters in the general pre-trained BERT model are updated according to the third loss function value to complete the fine-tuning training of the BERT model, and the fine-tuned BERT model is recorded as the semantic feature parsing model.

[0034] In one embodiment, the step of performing cross-modal alignment processing on the visual features, the temporal features, and the semantic features to generate a unified joint representation vector includes: The visual features, temporal features, and semantic features are input into a pre-defined cross-modal attention layer, which fuses these features to generate a unified joint representation vector. The core function of this cross-modal attention layer is to calculate the similarity matrix between features from different modalities, enabling features from one modality (such as temporal features) to "pay attention" to relevant information in other modalities (such as images or text). A specific implementation can be achieved through the attention mechanism shown in the following formula: In the formula, (Query) represents the query vector, which can be a time-series feature; (Key) represents a key vector, which can be an image feature or a text feature; (Value) represents a value vector, which can be image features or text features; The dimension representing the key vector is used as a scaling factor to prevent the gradient from becoming too small. Through this mechanism, the system can calculate the association weights between temporal features and image and text features, generating a unified joint representation vector that integrates information from multiple sources.

[0035] The cross-modal attention layer is obtained by training a joint model through the following steps: acquiring a power dispatch scenario dataset; wherein the power dispatch scenario dataset includes several power dispatch scenario training samples, each of which includes a set of equipment image training data, power grid load time-series training data, operation and maintenance log training data, and dispatch instruction training data acquired at the same time; the joint model includes the visual feature extraction model, the time-series feature encoding model, the semantic feature parsing model, and the cross-modal attention layer to be trained; and dividing the power dispatch scenario dataset into several batches of fourth training samples according to a preset batch size. Each batch of fourth training samples is sequentially input into the joint model for iterative training until a preset number of training rounds is reached. When the joint model receives a batch of fourth training samples, it calls the visual feature extraction model, the temporal feature encoding model, and the semantic feature parsing model respectively. Based on the equipment image training data, power grid load temporal training data, operation and maintenance log training data, and scheduling instruction training data in the current batch of fourth training samples, it generates training visual features, training temporal features, and training semantic features. These generated training visual features, training temporal features, and training semantic features are input into the cross-modal attention layer to generate a unified joint representation vector for training. Based on the cross-modal alignment loss function, a fourth loss function value is calculated according to the unified joint representation vector for training. The cross-modal alignment loss function is used to shorten the distance between each interconnected feature in the fourth training sample within the unified joint representation vector space and to widen the distance between unconnected features. Using a preset optimizer, the learnable parameters in the cross-modal attention layer are updated according to the fourth loss function value, completing the training of the cross-modal attention layer.

[0036] Through the above steps, the present invention successfully transforms multi-source heterogeneous data into a high-dimensional, information-intensive unified joint representation vector, which can comprehensively and deeply reflect the real-time integrated status of the power grid.

[0037] Step S3: Based on the unified joint representation vector, real-time electricity price data, and meteorological early warning data, generate an initial scheduling strategy through a preset reinforcement learning model; In a preferred embodiment, the step of generating an initial scheduling strategy based on a unified joint representation vector, real-time electricity price data, and meteorological early warning data using a preset reinforcement learning model includes: The unified joint representation vector, the real-time electricity price data, and the meteorological early warning data are combined to construct a current state vector; The current state vector is input into the reinforcement learning model so that the reinforcement learning model outputs an action containing the energy storage charging and discharging power setting value and the thermal power unit output adjustment command based on the current state vector, as the initial scheduling strategy; The reinforcement learning model was obtained through offline training in a power system simulation environment as follows: The reinforcement learning model is iteratively trained in the power system simulation environment until a preset number of training rounds are reached, thus completing the training of the reinforcement learning model. In each iteration, a simulation state vector from the power system simulation environment is obtained as the current simulation state. The current simulation state is input into the reinforcement learning model so that the reinforcement learning model generates a simulation action based on the current simulation state; The simulation action is submitted to the power system simulation environment, so that the power system simulation environment performs simulation execution according to the simulation action, and generates a simulation state at the next moment and a simulation execution result. Based on the simulation results, and using a multi-objective reward function, a reward value is calculated. An experience tuple is constructed based on the current simulation state, the simulation action, the reward value, and the simulation state at the next moment. A proximal policy optimization algorithm is used to update the learnable parameters in the reinforcement learning model based on the empirical tuple, thereby completing the update of the reinforcement learning model in this iteration.

[0038] In a preferred embodiment, the power system simulation environment is constructed in the following manner: Obtain a power system physical simulator; Acquire topology data of the power system, operating parameter data of power system equipment, and physical operating constraints of the power system; The topology data, the operating parameter data, and the physical operating constraints are configured into the power system physical simulator to construct the power system simulation environment.

[0039] Specifically, this step utilizes the comprehensive and integrated state awareness results generated in S2 to generate specific and executable scheduling instructions through a decision model aimed at optimizing power grid operation.

[0040] In a preferred embodiment, the step of generating an initial scheduling strategy based on a unified joint representation vector, real-time electricity price data, and meteorological early warning data using a preset reinforcement learning model includes: The unified joint representation vector, the real-time electricity price data, and the meteorological warning data are combined to construct a current state vector. This current state vector is a key input to the decision-making model in this embodiment of the invention. It not only includes a comprehensive perception of the physical state of the power grid formed by the multimodal data (visual, temporal, and textual) fused through step S2, but also supplements the key external factors driving economic dispatch (real-time electricity price) and safety dispatch (meteorological warning). The current state vector is input to the reinforcement learning model, so that the reinforcement learning model outputs an action containing the energy storage charging and discharging power setpoint and the thermal power unit output adjustment command based on the current state vector, as the initial dispatch strategy. The reinforcement learning model, as a pre-trained offline decision-making agent, aims to generate a set of optimal and directly executable dispatch command combinations that maximize future cumulative rewards based on a deep understanding of the current state vector, i.e., the initial dispatch strategy.

[0041] The reinforcement learning model is obtained through offline training in a power system simulation environment: the reinforcement learning model is iteratively trained in the power system simulation environment until a preset number of training rounds are reached, thus completing the training of the reinforcement learning model; In each iteration, a simulation state vector from the power system simulation environment is acquired as the current simulation state. This current simulation state is input into the reinforcement learning model, enabling the model to generate a simulation action based on the current simulation state. The simulation action is then submitted to the power system simulation environment, allowing the environment to perform a simulation execution based on the action, generating a simulation state for the next time step and a simulation execution result. The simulation execution result specifically reflects the changes in the power grid state brought about by the simulation action in the simulation environment, as well as the resulting economic benefits, power supply reliability indicators, and policy stability, among other feedbacks. Based on the simulation execution result and a multi-objective reward function, a reward value is calculated. In this embodiment, the multi-objective reward function aims to balance multiple conflicting objectives in scheduling, such as economy, reliability, and security. A specific multi-objective reward function R can be calculated using the following formula: In the formula, Represents the overall reward value; Represents economic benefits; This is an unreliable penalty item; This is a penalty for strategy fluctuations. The weighting coefficient for economic benefits; The weighting coefficient for the unreliable penalty term; The weighting coefficients are the weights of the policy fluctuation penalty term; these weighting coefficients are hyperparameters preset before training based on the scheduling objective.

[0042] The economic benefits can be further calculated using the following formula: In the formula, To obtain peak-valley price difference revenue through energy storage charging and discharging and other means; The cost of equipment wear and tear resulting from scheduling actions (such as unit adjustments and line losses).

[0043] The unreliability penalty term is used to quantify the degree to which the scheduling strategy violates grid security constraints, and it can be further calculated using the following formula: In the formula, This is to simulate the load reduction amount that appears in the execution results; This is the cumulative value of the voltage exceeding the limit that occurs in the simulation execution results; This is the cumulative value of the frequency exceeding the limit in the simulated execution results; The penalty weight for load reduction; The penalty weight for exceeding the voltage limit; The penalty weight for exceeding the frequency limit.

[0044] The policy fluctuation penalty term is used to suppress overly frequent or drastic adjustments in the model output, ensuring the stability of the policy. It can be further calculated using the following formula: In the formula, For the reinforcement learning model at the current simulation time Output simulation action (vector); For the reinforcement learning model at the previous simulation time... Output simulation action (vector); It is the square of the L2 norm (Euclidean norm), used to measure the magnitude of change in the vector between two consecutive actions.

[0045] Based on the current simulation state, the simulation action, the reward value, and the state at the next simulation moment, an empirical set M is formed. Using a proximal policy optimization (PPO) algorithm, the learnable parameters in the reinforcement learning model are updated according to the empirical set M to complete the update of the reinforcement learning model in this iteration. The proximal policy optimization algorithm ensures the stability of the training process by imposing constraints during policy updates, enabling the model to efficiently learn how to maximize the comprehensive reward value R in a complex simulation environment, thereby converging to an optimal scheduling policy network.

[0046] In a preferred embodiment, the power system simulation environment is constructed as follows: A power system physical simulator is acquired; topology data, operating parameter data of power system equipment, and physical operating constraints of the power system are acquired; the topology data includes the main wiring diagram of the power grid, line connection relationships, and power plant distribution. The operating parameter data includes the start-up and shutdown costs of thermal power units, ramp rates, rated capacity of energy storage devices, and charging and discharging efficiency. The physical operating constraints are crucial for ensuring the realism of the simulation and form the basis for compliance verification in subsequent step S4. Specifically, these constraints include the N-1 safety criteria for lines, load rate thresholds for transmission lines (e.g., load rate not exceeding 90%), upper and lower limits of the state of charge (SOC) of energy storage devices (e.g., SOC not lower than 20%), and the stability range of system frequency and voltage. The topology data, operating parameter data, and physical operating constraints are then configured into the power system physical simulator to construct the power system simulation environment.

[0047] Through the above steps, this invention obtains a reinforcement learning model that can deeply understand the multimodal state of the power grid and make dynamic decisions with the goal of maximizing the comprehensive benefits of the power grid, providing a core intelligent agent for generating high-quality initial scheduling strategies.

[0048] Step S4: Perform similarity matching calculation between the initial scheduling strategy and the preset historical best solution database to obtain the similarity matching result; perform compliance verification between the initial scheduling strategy and the preset physical constraint rule database to obtain the compliance verification result; determine the target scheduling strategy based on the similarity matching result and the compliance verification result. Specifically, this step is the core link in ensuring the reliability of AI decision-making in this embodiment of the invention. It uses a dual, orthogonal verification mechanism to automatically verify and filter the initial scheduling strategy generated in S3.

[0049] In a preferred embodiment, the step of performing similarity matching calculation between the initial scheduling strategy and a pre-set historical best solution database specifically refers to the following: the historical best solution database pre-stores a large number of scheduling cases that have been verified as safe and efficient in historical operation. These cases may include scheduling schemes under typical operating conditions and emergency response schemes under specific fault scenarios. The similarity matching calculation can use the cosine similarity algorithm, which measures the similarity between the initial scheduling strategy (as vector A) and the historical best solution stored in the database (as vector B).

[0050] In another preferred embodiment, the compliance verification of the initial scheduling strategy with a preset physical constraint rule database specifically refers to the following: the physical constraint rule database contains the safety red lines and physical limitations that the power system must adhere to during operation. These rules are consistent with the physical operating constraints used in step S3 to construct the simulation environment, and can be specifically set as: "Transmission line load rate not higher than 90%", "State of charge (SOC) of energy storage equipment not lower than 20%", and "The system must meet the N-1 safety criterion". The compliance verification process involves checking each instruction in the initial scheduling strategy (which includes energy storage charging and discharging power settings and thermal power unit output adjustment instructions) to determine whether it would lead to the violation of any of the above rules. If all rules are not violated, the compliance verification result is "passed"; otherwise, if any rule is violated, the result is "failed".

[0051] The system determines the target scheduling strategy based on similarity matching and compliance verification results. The decision logic is as follows: an initial scheduling strategy is considered safe and trustworthy only if its compliance verification result is "passed" and its similarity matching result is higher than a preset similarity threshold (e.g., 0.75). In this case, the system determines the initial scheduling strategy as the target scheduling strategy. If any of the above conditions are not met (i.e., compliance verification fails, or similarity is too low), the system can trigger an alarm, prompting the scheduler to manually review the data, or automatically revert to a preset safety baseline strategy to ensure absolute scheduling safety.

[0052] This dual verification mechanism effectively filters out scheduling schemes that may be generated by the AI ​​model, which may have security risks or deviate from optimal practices, thus ensuring the final reliability of scheduling decisions while achieving intelligence.

[0053] Step S5: Dispatch the power system according to the target scheduling strategy.

[0054] Specifically, this step includes: parsing the target scheduling strategy, which has been determined in S4 and has passed dual verification, into control instructions executable by power system automation equipment. The specific instructions included in the target scheduling strategy (such as energy storage charging and discharging power setpoints and thermal power unit output adjustment instructions) are sent to the power system's energy management system (EMS) or SCADA system, which then drive the corresponding physical equipment (such as energy storage units and generator sets) to perform the corresponding scheduling actions.

[0055] Furthermore, after the target scheduling strategy is executed, this method also includes a feedback optimization phase. The system collects the actual execution effect of the scheduling instructions in real time through Internet of Things (IoT) devices, such as the scheduling deviation rate and the actual load data of the lines. This actual execution effect is transmitted back as real-time feedback data for online updating and learning of the adapter parameters (such as LoRA adapter parameters) in the lightweight model. By constructing this closed-loop mechanism of "generation-verification-execution-feedback", the model can continuously self-optimize and constantly improve its adaptive ability to power grid dynamic changes and sudden faults.

[0056] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0057] like Figure 2 As shown, an embodiment of the present invention provides a power dispatching device that integrates multimodal data and dual verification, including: a power multimodal dynamic data acquisition module, a unified joint representation vector generation module, an initial dispatching strategy generation module, a target dispatching strategy determination module, and a power system dispatching module; The power multimodal dynamic data acquisition module is used to acquire multimodal dynamic data of the power system. The multimodal dynamic data includes equipment image data, power grid load time series data, operation and maintenance log data, dispatch instruction data, meteorological early warning data, and real-time electricity price data. The unified joint representation vector generation module is used to extract visual features from equipment image data to capture local abnormal areas and generate visual features; encode time-series load data to generate time-series features; parse semantic features from operation and maintenance log data and scheduling instruction data to generate semantic features; and perform cross-modal alignment processing on the visual features, the time-series features and the semantic features to generate a unified joint representation vector. The initial scheduling strategy generation module is used to generate an initial scheduling strategy based on a unified joint representation vector, real-time electricity price data, and meteorological early warning data, using a preset reinforcement learning model. The target scheduling strategy determination module is used to perform similarity matching calculation between the initial scheduling strategy and a preset historical optimal solution database to obtain a similarity matching result; and to perform compliance verification between the initial scheduling strategy and a preset physical constraint rule database to obtain a compliance verification result. The power system dispatching module is used to dispatch the power system according to the target dispatching strategy.

[0058] It should be noted that the embodiments of the device described above correspond to the embodiments of the present invention described above, and can realize the power dispatching method for fusing multimodal data and dual verification as described in any one of the above embodiments of the present invention. Furthermore, the embodiments of the device described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0059] Based on the above-described method embodiments of the present invention, a corresponding embodiment of an electronic device is provided.

[0060] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the power dispatching method integrating multimodal data and dual verification as described in any one of the present invention, or the processor executes the computer program to implement the functions of each module in the above-described device embodiments.

[0061] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0062] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0063] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0064] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0065] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments; Another embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute any of the above-described power dispatching methods for fusing multimodal data and dual verification.

[0066] The aforementioned storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0067] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0068] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A power dispatching method integrating multimodal data and dual verification, characterized in that, include: Acquire multimodal dynamic data of the power system, including equipment image data, grid load time-series data, operation and maintenance log data, dispatch instruction data, meteorological early warning data, and real-time electricity price data; Visual features are extracted from device image data to capture local abnormal areas and generate visual features; Time-series load data is encoded with time-series features to generate time-series features; semantic features are parsed from operation and maintenance log data and scheduling instruction data to generate semantic features; cross-modal alignment is performed on the visual features, the time-series features and the semantic features to generate a unified joint representation vector. Based on the unified joint representation vector, real-time electricity price data, and meteorological early warning data, an initial scheduling strategy is generated through a pre-set reinforcement learning model. The initial scheduling strategy is compared with a pre-set database of historical best solutions to calculate the similarity matching result. The initial scheduling strategy is compared with the preset physical constraint rule database for compliance verification, and the compliance verification results are obtained. Based on similarity matching results and compliance verification results, the target scheduling strategy is determined; The power system is scheduled according to the target scheduling strategy.

2. The power dispatching method integrating multimodal data and dual verification as described in claim 1, characterized in that, The step of extracting visual features from device image data to capture local abnormal regions and generating visual features includes: The device image data is input into a preset visual feature extraction model so that the visual feature extraction model generates visual features based on the device image data; The visual feature extraction model is generated by fine-tuning a pre-trained Vision Transformer model through the following training: Obtain a power system visual image training dataset; wherein, the power system visual image training dataset includes several image-text pairs, each image-text pair including an image of power system equipment and a descriptive text used to describe the operating status of the power system equipment; The power system visual image training dataset is divided into several batches of first training samples according to a preset batch size; The first training samples of each batch are sequentially input into the pre-trained Vision Transformer model for iterative fine-tuning until the preset number of training rounds is reached; The pre-trained Vision Transformer model generates corresponding image embedding vectors and text embedding vectors based on the power system equipment images and descriptive text in the current batch when it receives the first training samples of each batch. Based on a contrastive learning loss function, the first loss function value is calculated with the goal of maximizing the similarity between the matched image embedding vector and the text embedding vector. Using a preset optimizer, the learnable parameters in the pre-trained Vision Transformer model are updated according to the first loss function value to complete the training of the pre-trained Vision Transformer model, and the trained pre-trained Vision Transformer model is recorded as the visual feature extraction model.

3. The power dispatching method integrating multimodal data and dual verification as described in claim 2, characterized in that, The step of encoding time-series features of power grid load time-series data to generate time-series features includes: The power grid load time series data is input into a preset time series feature encoding model so that the time series feature encoding model generates time series features based on the power grid load time series data; The temporal feature encoding model is generated through the following training: Obtain a power grid time-series training dataset; wherein, the power grid time-series training dataset includes several time-series training samples, each of the time-series training samples includes power grid load data for a historical period, renewable energy output curve data, and corresponding future period real load fluctuation labels; The power grid time series training dataset is divided into several batches of second training samples according to a preset batch size; The second training samples of each batch are sequentially input into the temporal feature encoding model for iterative training until the preset number of training rounds is reached; The time-series feature coding model outputs predicted load fluctuations based on the historical grid load data and new energy output curve data of the current batch when it receives a batch of second training samples. Based on the mean squared error loss function, the second loss function value is calculated according to the predicted load fluctuation and the actual load fluctuation label corresponding to the time series training sample; Using a preset optimizer, the learnable parameters in the temporal feature coding model are updated according to the second loss function value, thereby completing the training of the temporal feature coding model.

4. The power dispatching method integrating multimodal data and dual verification as described in claim 3, characterized in that, The semantic feature parsing of operation and maintenance log data and scheduling instruction data to generate semantic features includes: The operation and maintenance log data and scheduling instruction data are input into a preset semantic feature parsing model, so that the semantic feature parsing model generates semantic features based on the operation and maintenance log data and scheduling instruction data; The semantic feature parsing model is obtained by fine-tuning a pre-trained BERT model as follows: Obtain a text training dataset in the power industry; wherein, the text training dataset in the power industry includes several text samples in the power industry, and the text samples in the power industry include historical operation and maintenance log data and scheduling instruction data; The power industry text training dataset is divided into several batches of third training samples according to a preset batch size; The third training samples from each batch are sequentially input into the general pre-trained BERT model for iterative fine-tuning until the preset number of training rounds is reached. The general pre-trained BERT model performs a masking language model task on the text samples in the current batch when it receives the third training sample in each batch, and outputs the prediction results of the masked words. Based on a cross-entropy loss function, the value of a third loss function is calculated according to the prediction result of the masked word and the corresponding real word in the text sample. Using a preset optimizer, the learnable parameters in the general pre-trained BERT model are updated according to the value of the third loss function to complete the fine-tuning training of the BERT model, and the fine-tuned BERT model is recorded as the semantic feature parsing model.

5. The power dispatching method integrating multimodal data and dual verification as described in claim 4, characterized in that, The step of performing cross-modal alignment processing on the visual features, the temporal features, and the semantic features to generate a unified joint representation vector includes: The visual features, temporal features, and semantic features are input into a preset cross-modal attention layer, so that the cross-modal attention layer fuses the visual features, temporal features, and semantic features to generate a unified joint representation vector; The cross-modal attention layer is obtained by training a joint model through the following steps: Obtain a power dispatch scenario dataset; wherein, the power dispatch scenario dataset includes several power dispatch scenario training samples, each power dispatch scenario training sample includes a set of equipment image training data, power grid load time series training data, operation and maintenance log training data and dispatch instruction training data acquired at the same time; the joint model includes the visual feature extraction model, the time series feature encoding model, the semantic feature parsing model and a cross-modal attention layer to be trained; The power dispatch scenario dataset is divided into several batches of fourth training samples according to a preset batch size; The fourth training samples from each batch are sequentially input into the joint model for iterative training until the preset number of training rounds is reached. When the joint model receives the fourth training sample in each batch, it calls the visual feature extraction model, the temporal feature encoding model and the semantic feature parsing model respectively, and generates training visual features, training temporal features and training semantic features based on the equipment image training data, power grid load temporal training data, operation and maintenance log training data and scheduling instruction training data in the current batch of the fourth training sample. The generated visual features, temporal features, and semantic features for training are input into the cross-modal attention layer to generate a unified joint representation vector for training. Based on the cross-modal alignment loss function, the fourth loss function value is calculated according to the unified joint representation vector used for training; wherein, the cross-modal alignment loss function is used to bring the distance between each interrelated feature in the fourth training sample in the unified joint representation vector space closer, and to push away the distance between unrelated features. Using a preset optimizer, the learnable parameters in the cross-modal attention layer are updated according to the fourth loss function value, thereby completing the training of the cross-modal attention layer.

6. The power dispatching method integrating multimodal data and dual verification as described in claim 5, characterized in that, The initial scheduling strategy is generated based on a unified joint representation vector, real-time electricity price data, and meteorological early warning data, using a pre-set reinforcement learning model. This includes: The unified joint representation vector, the real-time electricity price data, and the meteorological early warning data are combined to construct a current state vector; The current state vector is input into the reinforcement learning model so that the reinforcement learning model outputs an action containing the energy storage charging and discharging power setting value and the thermal power unit output adjustment command based on the current state vector, as the initial scheduling strategy; The reinforcement learning model was obtained through offline training in a power system simulation environment as follows: The reinforcement learning model is iteratively trained in the power system simulation environment until a preset number of training rounds are reached, thus completing the training of the reinforcement learning model. In each iteration, a simulation state vector from the power system simulation environment is obtained as the current simulation state. The current simulation state is input into the reinforcement learning model so that the reinforcement learning model generates a simulation action based on the current simulation state; The simulation action is submitted to the power system simulation environment, so that the power system simulation environment performs simulation execution according to the simulation action, and generates a simulation state at the next moment and a simulation execution result. Based on the simulation results, and using a multi-objective reward function, a reward value is calculated. An experience tuple is constructed based on the current simulation state, the simulation action, the reward value, and the simulation state at the next moment. A proximal policy optimization algorithm is used to update the learnable parameters in the reinforcement learning model based on the empirical tuple, thereby completing the update of the reinforcement learning model in this iteration.

7. The power dispatching method integrating multimodal data and dual verification as described in claim 6, characterized in that, The power system simulation environment is constructed using the following method: Obtain a power system physical simulator; Acquire topology data of the power system, operating parameter data of power system equipment, and physical operating constraints of the power system; The topology data, the operating parameter data, and the physical operating constraints are configured into the power system physical simulator to construct the power system simulation environment.

8. A power dispatching device integrating multimodal data and dual verification, characterized in that, include: The module includes a power multimodal dynamic data acquisition module, a unified joint representation vector generation module, an initial scheduling strategy generation module, a target scheduling strategy determination module, and a power system scheduling module. The power multimodal dynamic data acquisition module is used to acquire multimodal dynamic data of the power system. The multimodal dynamic data includes equipment image data, power grid load time series data, operation and maintenance log data, dispatch instruction data, meteorological early warning data, and real-time electricity price data. The unified joint representation vector generation module is used to extract visual features from device image data to capture local abnormal regions and generate visual features. Time-series load data is encoded with time-series features to generate time-series features; semantic features are parsed from operation and maintenance log data and scheduling instruction data to generate semantic features; cross-modal alignment is performed on the visual features, the time-series features and the semantic features to generate a unified joint representation vector. The initial scheduling strategy generation module is used to generate an initial scheduling strategy based on a unified joint representation vector, real-time electricity price data, and meteorological early warning data, using a preset reinforcement learning model. The target scheduling strategy determination module is used to perform similarity matching calculation between the initial scheduling strategy and the preset historical best solution database to obtain the similarity matching result; The initial scheduling strategy is compared with the preset physical constraint rule database for compliance verification, and the compliance verification results are obtained. The power system dispatching module is used to dispatch the power system according to the target dispatching strategy.

9. An electronic device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the power dispatching method for fusing multimodal data and dual verification as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the storage medium is located to perform the power dispatching method that integrates multimodal data and dual verification as described in any one of claims 1 to 7.

Images