Park load prediction method and device based on multi-modal data, and medium

Abstract

The embodiment of the invention discloses a park load prediction method and device based on multi-modal data and a medium, and relates to the technical field of micro-grids, and the method comprises the steps: collecting multi-modal micro-grid data through a multi-source heterogeneous sensor network disposed at a power distribution node, the multi-mode micro-grid data comprises electrical measurement data, environment monitoring data and equipment state data; constructing a dynamic space-time incidence matrix according to the multi-modal micro-grid data, and performing feature extraction based on the dynamic space-time incidence matrix through a space-time diagram convolutional network deployed at an edge calculation node to determine a space-time feature vector; and inputting the spatio-temporal feature vector into a time sequence fusion prediction model, outputting a load prediction result, and generating an optical storage cooperative scheduling instruction according to the load prediction result, the load prediction result including a load prediction value and a confidence interval.

Description

Technical Field

[0001] This specification relates to the technical field of microgrids, and particularly to a method, device, and medium for predicting park loads based on multi-modal data. Background Art

[0002] Accurate load forecasting can help parks formulate reasonable energy procurement plans, avoid relying too much on the main grid for power supply during peak electricity price periods, and thus reduce electricity costs. At the same time, combined with renewable energy output forecasting, the charging and discharging strategies of energy storage systems can be optimized to maximize the economic benefits of storing energy during low-price periods and discharging during high-price periods.

[0003] In the prior art, park microgrid load forecasting mainly relies on statistical modeling methods of historical load curves (such as ARIMA, linear regression) and deep learning models based on LSTM. However, with the high proportion of renewable energy access and the frequent changes in dynamic topological structures, traditional methods have obvious technical drawbacks. First, traditional forecasting methods mainly rely on the statistical characteristics of historical load curves (such as ARIMA, linear regression), and it is difficult to cope with the strong randomness and minute-level fluctuations of distributed energy sources such as photovoltaic and wind power, resulting in high prediction errors. Especially under complex working conditions such as sudden weather changes or equipment commissioning and decommissioning, the prediction deviation increases exponentially. Although existing deep learning solutions introduce time series models such as LSTM and CNN, their spatial feature extraction relies on graph neural networks with fixed topologies. When new photovoltaic arrays or load nodes are added to the park, the model needs to be retrained and cannot adapt to dynamic topological changes. In addition, the cloud-based centralized computing architecture leads to response delays exceeding real-time control requirements, and the limited memory resources of edge devices are difficult to carry complex model deployments, seriously restricting the applicability in industrial scenarios.

[0004] Therefore, in scenarios with high proportions of renewable energy access and frequent changes in dynamic topological structures, traditional load forecasting methods have large prediction errors, cannot adapt to dynamic topological changes within the park, and cannot meet real-time scheduling requirements. Summary of the Invention

[0005] One or more embodiments of this specification provide a method, device, and medium for predicting park loads based on multi-modal data, which are used to solve the following technical problems: In scenarios with high proportions of renewable energy access and frequent changes in dynamic topological structures, traditional load forecasting methods have large prediction errors, cannot adapt to dynamic topological changes within the park, and cannot meet real-time scheduling requirements.

[0006] One or more embodiments of this specification adopt the following technical solutions:

[0007] One or more embodiments of this specification provide a method for predicting park load based on multimodal data. The method includes: collecting multimodal microgrid data through a multi-source heterogeneous sensor network deployed at distribution nodes, where the multimodal microgrid data includes electrical measurement data, environmental monitoring data, and equipment status data; constructing a dynamic spatio-temporal correlation matrix based on the multimodal microgrid data, and performing feature extraction based on the dynamic spatio-temporal correlation matrix through a spatio-temporal graph convolutional network deployed at an edge computing node to determine a spatio-temporal feature vector; inputting the spatio-temporal feature vector into a temporal fusion prediction model to output a load prediction result, and generating a coordinated dispatching instruction for photovoltaic energy storage based on the load prediction result, where the load prediction result includes a load prediction value and a confidence interval.

[0008] One or more embodiments of this specification provide a park load prediction device based on multimodal data, including:

[0009] At least one processor; and,

[0010] A memory communicatively connected to the at least one processor; where,

[0011] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above method.

[0012] A non-volatile computer storage medium provided by one or more embodiments of this specification stores computer-executable instructions, and the computer-executable instructions are set to: execute the above method.

[0013] The above at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: Through the technical solution of the embodiments of this specification, traditional load forecasting methods mainly rely on the statistical characteristics of historical load curves. In the embodiments of this specification, by collecting multi-modal microgrid data, various data complement each other, enabling the model to more comprehensively consider various factors affecting the load, thereby effectively reducing the forecasting error, improving the accuracy of load forecasting, and avoiding the problem that the forecasting deviation of traditional methods grows exponentially under complex working conditions; Regarding the problem that in existing deep learning solutions, the extraction of spatial features relies on graph neural networks with fixed topologies and cannot adapt to the dynamic changes of topologies, in the embodiments of this specification, after collecting multi-modal microgrid data, abnormal events such as equipment commissioning and decommissioning can be monitored in real time through equipment status data. When the topology structure changes, it can trigger event-driven edge weight updates and perform local updates on the adjacency matrix, thereby constructing a dynamic spatio-temporal correlation matrix, enabling the spatio-temporal graph convolutional network based on this matrix to adapt to the dynamic changes of the topology structure without retraining the model, and improving the adaptability of the model to the dynamic changes of the campus microgrid; In addition, the cloud centralized computing architecture results in response delays exceeding real-time control requirements, and edge computing nodes are difficult to carry complex model deployments due to limited memory resources. In the embodiments of this specification, by collecting multi-modal microgrid data and processing it at the edge computing nodes, on the one hand, it can reduce the time and communication cost of data transmission to the cloud and reduce the response delay; on the other hand, the collection and processing of multi-modal data can complete some operations such as feature extraction at the edge computing nodes, reducing the pressure on the memory resources of edge devices, enabling the edge computing nodes to better carry the model operation and meet the requirements of real-time scheduling; At the same time, the load forecasting results based on multi-modal data can generate more accurate photovoltaic-storage collaborative scheduling instructions, such as generating hierarchical scheduling instructions according to the deviation rate between the load forecasting value and the real-time load monitoring value, realizing local load transfer, cloud photovoltaic-storage collaborative optimization, and protective load shedding operations, etc., improving energy utilization efficiency and system operation stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments recorded in this specification. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In the drawings:

[0015] Figure 1 It is a schematic flowchart of a method for predicting campus load based on multi-modal data provided by the embodiments of this specification;

[0016] Figure 2Schematic diagram of the structure of a campus load prediction device based on multimodal data provided by the embodiments of this specification. Detailed implementation manners

[0017] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments of this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this specification.

[0018] The embodiments of this specification provide a campus load prediction method based on multimodal data. It should be noted that the execution subject in the embodiments of this specification can be a server or any device with data processing capabilities. Figure 1 Schematic flowchart of a campus load prediction method based on multimodal data provided by the embodiments of this specification, as Figure 1 shown, mainly including the following steps:

[0019] Step S101, collect multimodal microgrid data through a multi-source heterogeneous sensor network deployed at the power distribution node.

[0020] In an embodiment of this specification, a multi-source heterogeneous sensing network is deployed at the power distribution node to collect multimodal microgrid data. The multimodal microgrid data includes electrical measurement data, environmental monitoring data, and equipment status data. The electrical measurement data is collected through a broadband synchronous sampling device deployed at the power distribution node, the environmental monitoring data is collected through a temperature and humidity sensor and a light intensity meter, and the equipment status data is collected through a vibration sensor and an infrared thermal imaging device. Among them, for electrical measurement, a broadband synchronous acquisition device with a sampling rate not less than 256 points / cycle is used, which can accurately capture the transient characteristics of the load. The environmental monitoring is equipped with a temperature and humidity sensor with an accuracy of ±0.2 ℃ and a high-dynamic range light meter with a range of 0 - 2000 W / m 2 to ensure the accurate acquisition of environmental parameters. The equipment status monitoring innovatively integrates a vibration sensor with a frequency response range of 10 Hz - 1 kHz and an infrared thermal imager with a resolution of 640×480, realizing all-round status perception of key equipment such as transformers and inverters. The protocol-compatible data acquisition gateway supports multi-protocol conversion such as Modbus TCPJP and IEC61850, and has an 8GB data cache built-in to ensure data integrity during network interruptions.

[0021] Preprocess the multi-modal microgrid data, perform waveform integrity verification on the electrical measurement data, perform rationality verification on the voltage parameters in the electrical measurement data based on the rated voltage range, and screen out abnormal data segments beyond the preset fluctuation range; compensate for the missing electrical measurement data using the weighted interpolation algorithm of the same type of data from adjacent nodes, and the weight value is dynamically adjusted according to the electrical distance between nodes; perform abnormal correction on the temperature parameters in the environmental monitoring data based on the equipment operation status, and associate the air conditioner start-stop events to correct the temperature acquisition deviation. Map the multi-source data to a unified time base through the timestamp alignment protocol.

[0022] Through the above technical solution, by collecting multi-modal microgrid data, the electrical measurement data therein can reflect the current operation status of the microgrid, such as voltage, current and other information, providing direct electrical operation parameters for load forecasting; the environmental monitoring data can capture the impact of environmental factors such as weather on the power generation of distributed energy (such as photovoltaic and wind power) and the load in the park, such as temperature, light intensity, etc. These factors will significantly affect the output of distributed energy and the electricity consumption behavior of users; the equipment status data helps to understand the operation status of the equipment and timely detect the impact of abnormal events such as equipment commissioning and decommissioning on the load. The multiple data complement each other, enabling the model to more comprehensively consider various factors affecting the load, thereby effectively reducing the prediction error, improving the accuracy of load forecasting, and avoiding the problem that the prediction deviation of traditional methods grows exponentially under complex working conditions.

[0023] Step S102, according to the multi-modal microgrid data, construct a dynamic spatio-temporal correlation matrix, and based on the spatio-temporal graph convolutional network deployed on the edge computing node, perform feature extraction based on the dynamic spatio-temporal correlation matrix to determine the spatio-temporal feature vector.

[0024] Construct a dynamic spatio-temporal correlation matrix according to the multi-modal microgrid data, specifically including: taking the distribution transformers, photovoltaic arrays and energy storage converters in the park microgrid as graph nodes, and calculating the equivalent impedance parameters between nodes according to the electrical measurement data in the multi-modal microgrid data to generate an initial adjacency matrix; perform abnormal event monitoring according to the equipment status data in the multi-modal microgrid data, where the abnormal event includes equipment commissioning and decommissioning events or load power mutation events; when an abnormal event is detected, trigger event-driven edge weight update to update the local adjacency matrix of the initial adjacency matrix and construct the dynamic spatio-temporal correlation matrix.

[0025] In one embodiment of this specification, key devices such as distribution transformers, photovoltaic arrays, and energy storage converters are used as graph nodes. The multimodal microgrid data corresponding to each node includes electrical measurement data, environmental monitoring data, and device status data. The electrical measurement data includes parameters such as voltage waveforms, effective current values, and harmonic components; the device status data is obtained through vibration sensors, infrared thermal imagers, etc., and is used to monitor the commissioning and decommissioning status of devices and abnormal operating conditions. All data is standardized through a gateway that supports multi-protocol conversion to ensure unified data formats and timestamps. Based on the voltage phase difference and effective current value in the electrical measurement data, the equivalent impedance between nodes is calculated to generate an initial adjacency matrix. The device status data is used to monitor abnormal events in real time, including device commissioning and decommissioning and sudden changes in load power. The opening and closing signals of the circuit breaker and the vibration spectrum characteristics of the device are analyzed to identify the commissioning and decommissioning status of devices such as photovoltaic arrays and energy storage converters. The change rate of the load power is statistically analyzed through a sliding window, and when the change rate continuously exceeds the threshold, it is determined as an effective mutation event. After an abnormal event is detected, the affected nodes and associated edges are marked, triggering an update of the local adjacency matrix. For the nodes and edges affected by the abnormal event, the adjacency matrix is dynamically adjusted. If a device is taken out of service, all the edge weights associated with it are removed; if a device is put into service, the equivalent impedance between it and the adjacent nodes is recalculated and new edges are added; for the mutated nodes, the equivalent impedance is recalculated, the edge weights between it and the adjacent nodes are adjusted, and the update is extended to the second-order neighborhood nodes; an incremental calculation strategy is adopted to only update the affected local area, avoiding the computational overhead of full matrix reconstruction.

[0026] Integrate the output of the above steps to generate a dynamic spatio-temporal correlation matrix, which is used as the input of the spatio-temporal graph convolutional network to extract the spatial features of the nodes and fuse with the time series change rate of the environmental data, ultimately supporting load forecasting and optimal scheduling. The system continuously optimizes the matrix update strategy through a cloud-edge collaboration mechanism to ensure real-time adaptability to the dynamic topology of the microgrid.

[0027] According to the electrical measurement data in the multimodal microgrid data, calculate the equivalent impedance parameters between nodes to generate an initial adjacency matrix, specifically including: extracting the real and imaginary parts of the fundamental voltage component in the electrical measurement data, calculating the absolute value of the phase difference between nodes, and deriving the equivalent impedance between nodes according to Ohm's law; linearly combining the absolute value of the phase difference between nodes and the equivalent impedance between nodes according to a preset weight coefficient to generate a normalized distance index; marking valid edges according to the physical connection status between nodes to eliminate redundant associated edges between non-directly connected nodes; through an exponential mapping function, converting the normalized distance index into the edge weight of the adjacency matrix to generate an initial adjacency matrix, where the edge weight and the normalized distance index in the exponential mapping function are negatively correlated.

[0028] In one embodiment of this specification, voltage waveform signals are collected in real time by broadband power quality monitoring devices deployed at each node. Fourier transform is performed on the original voltage waveform to separate the fundamental frequency component (such as 50Hz power frequency), and the real and imaginary part values are extracted, which respectively correspond to the cosine term and sine term amplitudes of the fundamental component. Based on the numerical relationship between the real part and the imaginary part, the fundamental voltage phase angle of each node is calculated. For any two nodes, the absolute value of the difference between their phase angles is taken to obtain the voltage phase difference between the nodes. This phase difference reflects the synchronism of the voltage waveforms between the nodes, and the smaller the phase difference, the closer the electrical coupling. Based on Ohm's law, the equivalent impedance is calculated using the voltage difference between the nodes and the effective current value. Specifically, by measuring the effective current value of the line connecting the two nodes and combining the effective voltage difference corresponding to the voltage phase difference, the modulus value of the impedance is deduced. The equivalent impedance characterizes the degree of obstruction to energy transfer between the nodes, and the smaller the impedance value, the stronger the electrical connection.

[0029] The absolute value of the phase difference and the equivalent impedance are used as input features and linearly combined according to the preset weight coefficients. The weight coefficients are dynamically adjusted according to the type of microgrid (such as distribution network, industrial microgrid). For example, the weight of the phase difference is increased in a photovoltaic-dominated microgrid, and the impedance parameter is emphasized in a scenario with dense energy storage. The combined result is processed by maximum-minimum normalization and mapped to a unified dimension range (such as 0 to 1) to generate a normalized distance index. This index comprehensively characterizes the degree of electrical isolation between the nodes, and the larger the index value, the farther the electrical distance. According to the distribution network topology diagram provided by the SCADA system, node pairs with direct physical connections are marked. For non-directly connected nodes, such as nodes connected across transformers or through multi-stage switches, even if their normalized distance index is low, they are still regarded as redundant associated edges and removed. This step ensures that the adjacency matrix only retains the edges with actual electrical paths, avoiding noise interference introduced by misjudgment of electrical coupling. An exponential mapping function is designed to convert the normalized distance index into the edge weight of the adjacency matrix. This function satisfies the monotonically decreasing property, that is, when the normalized distance index increases, the edge weight decays according to an exponential law. In specific implementation, by adjusting the decay rate parameter of the exponential function, the sensitivity of the edge weight to distance changes is controlled. The finally generated edge weight ranges from 0 to 1, and the closer the weight is to 1, the stronger the correlation between the nodes, and vice versa. This adjacency matrix is used as the input of the spatio-temporal graph convolutional network to support subsequent spatial feature extraction.

[0030] Through the above technical solution, by extracting the real and imaginary parts of the fundamental voltage component to calculate the phase difference and combining Ohm's law to deduce the equivalent impedance, the electrical connection relationship and the degree of mutual influence among the nodes in the microgrid can be accurately reflected; the absolute value of the phase difference between nodes and the equivalent impedance are linearly combined to generate a normalized distance index, and the effective edges are marked according to the physical connection state between nodes, and the redundant associated edges between non-directly connected nodes are removed, which can remove unnecessary associated information, reduce the data complexity, improve the calculation efficiency and accuracy of the model, enable the model to focus on the relationship between nodes with actual electrical connections, and avoid misjudgment and calculation deviation caused by redundant information; the device status data is used for abnormal event monitoring. When a device connection or disconnection event or a load power mutation event occurs, an event-driven edge weight update is triggered to locally update the initial adjacency matrix. The dynamic spatio-temporal correlation matrix can reflect the changes in the topological structure and operating state of the microgrid in real time, quickly adapt to the grid changes, ensure that the load prediction is not affected by the topological structure changes, and effectively solve the problem that the traditional method cannot adapt to dynamic topological changes; the normalized distance index is converted into the edge weight of the adjacency matrix through an exponential mapping function, and the edge weight is negatively correlated with the normalized distance index. The closer the distance and the closer the electrical connection between nodes, the higher the edge weight, which more reasonably reflects the actual association degree between nodes, provides a more accurate basis for subsequent feature extraction and analysis based on models such as graph convolutional networks, and helps to improve the accuracy and reliability of load prediction.

[0031] After locally updating the initial adjacency matrix, the method further includes: recording the last update timestamp of the edge weight of each adjacency matrix edge, and calculating the update time interval between the current time and the last update timestamp; when the update time interval of a specified adjacency matrix edge is greater than a preset non-update duration, counting the historical adjacency matrix update frequency corresponding to the specified adjacency matrix edge; based on the historical adjacency matrix update frequency, applying an exponential decay factor to the edge weight of the specified adjacency matrix edge to determine a decaying edge weight; when the decaying edge weight is lower than a preset failure threshold, performing a removal operation on the specified adjacency matrix edge to generate an updated adjacency matrix.

[0032] In one embodiment of this specification, after completing the event-driven local adjacency matrix update, it is necessary to perform timeliness maintenance on the adjacency matrix to ensure that it dynamically reflects the real-time topological changes of the microgrid. Each time the local adjacency matrix update is executed, the current system time is recorded as the last update timestamp for the affected edges. The timestamp uses a unified clock source, and the accuracy needs to meet the millisecond-level requirement. For the edges that did not participate in this update, their original timestamps remain unchanged. The timestamp data is stored in the form of key-value pairs, where the key is the node pair identifier of the edge (such as "Node A - Node B"), and the value is the timestamp value. Based on the current system time and the last update timestamp of the edge weight, calculate the unupdated duration (i.e., time interval) of each edge. When the unupdated duration of an edge exceeds the preset threshold, it is determined that the edge enters the timeliness attenuation processing flow. The preset threshold is set according to the typical frequency of microgrid topological changes. For example, in a microgrid with frequent reconstruction, a shorter threshold is set to accelerate the elimination of stale associations. For the edges that enter the timeliness attenuation process, trace their historical update records and count the number of updates within the preset time window. For example, count the number of times this edge has been updated due to equipment connection / disconnection or load mutation in the past week, and calculate its average update frequency. The higher the update frequency, the stronger the dynamic association between the corresponding nodes of the edge, and the attenuation rate needs to be reduced to retain the effective association. Dynamically determine the attenuation factor according to the historical update frequency. For edges with high-frequency updates, use a smaller attenuation factor to slow down the weight attenuation speed and retain the recently active associations; for edges with low-frequency updates, use a larger attenuation factor to accelerate the weight to zero and quickly eliminate the invalid associations that have not been updated for a long time. The specific value of the attenuation factor is determined through a preset mapping relationship. For example, after normalizing the update frequency to the 0-1 interval, it is converted into an attenuation factor according to a linear or non-linear function.

[0033] Perform an exponential decay operation on the weight of the edge to be attenuated to obtain the attenuated weight value. The decay formula satisfies the characteristic that the weight decreases exponentially with the time interval. When the attenuated weight is lower than the preset failure threshold, it is determined that the edge has lost its effective association, and it is removed from the adjacency matrix. The failure threshold is usually set as a decimal close to zero to avoid accidentally deleting weak but effective associations. After completing the timeliness processing of all edges, generate the updated adjacency matrix. Perform topological connectivity verification on the updated matrix to ensure that the connectivity of key nodes (such as main transformers, energy storage PCS) is not damaged due to edge removal. If an abnormal disconnection is found, trigger the manual review process and record the log to prevent misoperations caused by automated processing.

[0034] Through the above technical solutions, by recording the last update timestamp of the edge weights and calculating the update time interval, the activity level of each edge in the network can be tracked in real time. When the update time interval is greater than the preset non-update duration, it indicates that the relationship between the nodes corresponding to this edge may not have changed for a long time, and its relevance to the real-time operation state of the system decreases. This enables the model to promptly capture the dynamic changes in the relationships between nodes in the system and more accurately reflect the real-time operation status of the microgrid. For the edges in the adjacency matrix that have not been updated for a long time, an exponential decay factor is applied based on their historical update frequency to determine the decaying edge weights, which can effectively avoid the problem of information obsolescence caused by long-term data non-update. When the decaying edge weight is lower than the preset failure threshold, a removal operation is executed to optimize the network structure of the dynamic spatio-temporal association matrix. Removing the edges that have been inactive for a long time and whose weights have decayed to a very low level can reduce redundant information in the model, lower the computational complexity, improve the operation efficiency and accuracy of the model, and also enable the model to focus on more important and active node relationships, better mining the key electrical connections and interactions in the microgrid. Through this dynamic edge weight management and edge removal mechanism, the constructed dynamic spatio-temporal association matrix can be made more robust and reliable. During the operation of the microgrid, the device status and electrical measurement data are constantly changing, and this method can adaptively adjust the matrix structure and promptly eliminate invalid or unreliable edge information that may be caused by equipment failures, communication interruptions, etc.

[0035] Through the spatio-temporal graph convolutional network deployed on the edge computing nodes, feature extraction is performed based on this dynamic spatio-temporal association matrix to determine the spatio-temporal feature vector, which specifically includes: performing polynomial approximation decomposition on the dynamic spatio-temporal association matrix to generate the basis function coefficients of the graph convolutional kernel and inputting them into the spatio-temporal graph convolutional network; determining the multi-dimensional feature vector according to the electrical measurement data, environmental monitoring data, and device status data, where the multi-dimensional feature vector includes the node voltage phase difference, environmental temperature change rate, and device vibration spectrum characteristics; inputting the multi-dimensional feature vector into the cross-modal attention layer to fuse the time series change rate corresponding to the environmental monitoring data to generate the spatio-temporal feature vector.

[0036] In one embodiment of this specification, a spatio-temporal graph convolutional network (ST-GCN) is deployed at the edge computing node. To extract spatio-temporal feature vectors based on the dynamic spatio-temporal correlation matrix, efficient feature expression needs to be achieved through multi-source data fusion and lightweight computing. The dynamic spatio-temporal correlation matrix is input into the preprocessing module of the edge computing node, and spectral decomposition is performed to analyze its graph structure characteristics. The adjacency matrix is approximated by low-order orthogonal polynomial basis functions (such as Chebyshev polynomials) to generate a set of basis function coefficients. These coefficients characterize the action range and intensity of the graph convolution kernel in the local neighborhood, replacing the complex Laplacian matrix decomposition operation in traditional graph convolution. During the decomposition process, the polynomial order is dynamically adjusted according to the spectral radius of the adjacency matrix to ensure that key spatial correlation information is retained while reducing the computational complexity. The finally generated basis function coefficients are combined with the node feature vectors to drive the feature propagation in the spatio-temporal graph convolutional layer.

[0037] The real-time data stream is obtained from the local cache of the edge node, and multi-modal inputs are processed in parallel. Fourier transform is performed on the voltage waveform collected by the broadband sampling device to extract the fundamental wave phase angle, and the absolute value of the phase difference between nodes is calculated to characterize the electrical synchronization. Sliding window difference calculation is performed on the time series data of the temperature and humidity sensors to obtain the temperature change rate, reflecting the driving effect of the external environment on the load. Fast Fourier transform is performed on the time domain signal of the vibration sensor to extract the main frequency band energy distribution and generate the equipment health status index. After the above features are normalized, they are concatenated into a multi-dimensional feature vector and input as the node attributes of the spatio-temporal graph convolutional network. The basis function coefficients and the multi-dimensional feature vector are input into the spatio-temporal graph convolutional layer to perform feature aggregation in the spatial domain. According to the edge weights of the dynamic adjacency matrix, the multi-dimensional features of adjacent nodes are aggregated. For example, higher weights are assigned to nodes with closely coupled electricity, strengthening the local spatial correlation. The features of high-degree nodes (such as distribution transformers) are normalized to avoid feature over-smoothing caused by too many connected edges. A gated recurrent unit (GRU) is stacked in the convolutional output layer to capture the temporal evolution law of the node state and suppress noise interference.

[0038] The spatial features output by the spatio-temporal graph convolution and the time series change rate of the environmental data are input into the cross-modal attention layer to align the spatial features (such as node load distribution) and the environmental temporal features (such as the temperature change rate curve) along the time axis to ensure data synchronization. The cosine similarity is used to measure the correlation strength between the electrical features and the environmental features, generating a dynamic attention weight matrix. High weights focus on strongly correlated feature combinations (such as high temperature periods and cooling loads). The Sigmoid function is used to generate the gating value, applying random dropout (Dropout) to the low-correlation feature channels and retaining the high-weight features for weighted fusion. The fused features are reduced in dimension through a fully connected layer, and the spatio-temporal feature vector is output, synchronously containing spatial topological correlation and environmental temporal driving information.

[0039] Through the above technical solutions, the polynomial approximation decomposition of the dynamic spatio-temporal correlation matrix is used to generate the basis function coefficients of the graph convolution kernel, which can transform complex graph structure data into a form suitable for convolution operations, enabling the spatio-temporal graph convolution network to better handle the complex relationships and dynamic changes between nodes and extract effective spatial features; determining the multi-dimensional feature vector according to electrical measurement data, environmental monitoring data, and equipment status data, and fusing data of different types and modalities together; not only considering the electrical information in the microgrid, such as the node voltage phase difference, but also incorporating non-electrical information such as the environmental temperature change rate and the equipment vibration spectrum characteristics, which can more comprehensively reflect the operating state of the microgrid and avoid the information limitation brought by a single data type; by fusing the time series change rate corresponding to the environmental monitoring data through the cross-modal attention layer, it can automatically learn the importance of different modal data at different time steps, highlight the key features that have a greater impact on the operating state of the microgrid, and suppress unimportant information, so that the generated spatio-temporal feature vector is more representative and discriminative, which helps to perform the prediction task more accurately; deployed on the edge computing node, it can perform real-time data processing and feature extraction near the data source, reducing the time and bandwidth consumption of data transmission to the cloud or the central server, greatly improving the processing speed, and providing timely and accurate information support for the real-time monitoring and decision-making of the microgrid.

[0040] Step S103, input the spatio-temporal feature vector into the time series fusion prediction model, and output the load prediction result, so as to generate the optical storage collaborative scheduling instruction according to the load prediction result.

[0041] Among them, the load prediction result includes the load prediction value and the confidence interval.

[0042] Inputting the spatio-temporal feature vector into the time series fusion prediction model and outputting the load prediction result specifically includes: inputting the spatio-temporal feature vector into the time series fusion prediction model, parsing the feature importance of different time granularities through the quantile attention layer to generate the prediction interval; setting a physical constraint term in the time series fusion prediction model to verify whether the initial prediction value conforms to the node power conservation law; if the initial prediction value does not conform to the node power conservation law, substituting the initial prediction value into the power flow equation to calculate the node voltage deviation; according to the node voltage deviation, adjusting the prediction value distribution of the associated load nodes until the voltage of all nodes is within the safe operating range and the power error rate is lower than the preset threshold.

[0043] In one embodiment of this specification, the process of inputting spatio-temporal feature vectors into a time-series fusion prediction model to generate load prediction results needs to integrate data-driven prediction and physical law constraints to ensure that the output results have both high accuracy and engineering feasibility. After the spatio-temporal feature vectors are input into the time-series fusion prediction model, the importance of the time dimension of the features is first analyzed through a quantile attention layer. This layer uses a multi-head attention mechanism to parallelly generate multiple query vectors for different confidence levels (such as P10, P50, P90). Each query vector is matched with the time-series pattern in the spatio-temporal features to calculate the attention weights for each time granularity. High-weight feature segments (such as the cooling load features during a sudden temperature rise period) contribute more to the prediction interval of the corresponding quantile. Finally, a prediction interval covering the probability distribution is output, where the P10 - P90 interval reflects the possible fluctuation range of future loads, providing a quantitative basis for risk dispatching.

[0044] In the model training stage, a power conservation constraint term is introduced into the loss function. Based on Kirchhoff's current law, this constraint term calculates the balance error between the input power, output power, and losses of each node, and is superimposed on the prediction loss as an additional penalty term to force the model to learn a prediction pattern that conforms to the circuit law. In the prediction stage, online verification is performed on the initial prediction values. All nodes are traversed to check whether the algebraic sum of the predicted powers satisfies the balance condition. If the power deviation of a node exceeds the allowable threshold, it is marked as a violation prediction, triggering a subsequent correction process. For the violation prediction results, they are substituted into a preset simplified power flow equation to calculate the node voltage. According to the deviation direction of the power flow calculation result from the rated voltage (such as the voltage being too high or too low), the predicted values of the load nodes that need to be adjusted are deduced in reverse. The correction strategy uses the principle of gradient descent to fine-tune the distribution of the predicted values of the associated load nodes along the direction of decreasing voltage deviation, while keeping the predicted values of other nodes relatively stable. After each correction, the power flow calculation and power balance verification are performed again until the voltage of all nodes is within the safe operating range and the power error rate is lower than the preset threshold. This process is implemented through a lightweight iterative algorithm of edge computing nodes to ensure that the correction delay is controllable.

[0045] Dynamically optimize the voltage safety range and power error threshold according to historical operation data. When the microgrid is in a high renewable energy penetration condition (such as the photovoltaic output exceeding 50%), the allowable voltage fluctuation range is appropriately relaxed to avoid prediction distortion caused by excessive correction; during peak load periods, the threshold is tightened to improve power supply quality. The threshold adjustment signal is sent by the cloud collaborative management platform, and the edge node updates the local correction strategy parameters in real time after receiving it.

[0046] Through the above technical solutions, by parsing the feature importance of different time granularities through the quantile attention layer, the changing rules of load data at different time scales can be captured, and a prediction interval can be generated, so as to more comprehensively and accurately reflect the changing trend of the load, and improve the accuracy and reliability of the prediction; a physical constraint term is set in the time series fusion prediction model to verify whether the initial prediction value conforms to the node power conservation law, which ensures that the prediction result is physically reasonable and conforms to the actual operation law of the microgrid, and improves the credibility and practicability of the prediction; when the initial prediction value does not conform to the node power conservation law, by substituting the initial prediction value into the power flow equation to calculate the node voltage deviation, and adjusting the prediction value distribution of the associated load nodes according to the deviation, all node voltages can be kept within the safe operation interval and the power error rate can be lower than the preset threshold, maintaining the voltage stability and power balance of the microgrid, and avoiding problems such as voltage fluctuations and power imbalances caused by inaccurate load prediction.

[0047] Generate a coordinated dispatching instruction for photovoltaic energy storage according to the load prediction result, which specifically includes: parsing the load prediction value, and calculating the deviation rate between the load prediction value and the real-time load monitoring value corresponding to the electrical measurement data; generating a hierarchical dispatching instruction according to the deviation rate, where the hierarchical dispatching instruction includes local load transfer, cloud photovoltaic energy storage collaborative optimization, and protective load shedding operation.

[0048] In an embodiment of the present specification, the load prediction value and the real-time load monitoring value collected by the broadband power quality monitoring device are obtained in real time at the edge computing node, and a sliding window statistics is performed on the two. Calculate the absolute deviation rate between the predicted value and the measured value within the window, and analyze its distribution characteristics (such as mean, variance, and continuous overrun duration). Divide the deviation rate into three levels according to the preset dynamic threshold strategy.

[0049] The first-level deviation is a low risk. At this time, the deviation rate is within the allowable fluctuation range, indicating that the prediction error is within the system tolerance, and the local load transfer strategy is triggered; the second-level deviation is a low to medium risk. At this time, the deviation rate continuously exceeds the safety threshold but does not reach the emergency level, and the cloud photovoltaic energy storage collaborative optimization needs to be started; the third-level deviation is a high risk. At this time, the deviation rate rises sharply and is accompanied by the risk of voltage / frequency overlimit, and the protective load shedding operation is immediately executed. The dynamic threshold is adaptively adjusted according to the historical error distribution, topological state, and environmental factors (such as sudden temperature changes) to avoid misjudgment caused by a fixed threshold.

[0050] For the first-level deviation, the edge node generates a load transfer plan based on a preset economic rule, retrieves the list of interruptible loads (such as non-critical production equipment and air-conditioning group control systems) from the load database, and calculates its maximum transferable power and time window; according to the time-of-use electricity price signal and user agreement, estimates the compensation cost of load transfer, and preferentially selects the load combination with the optimal economy; sends a transfer instruction to the target load controller through an industrial Internet of Things protocol (such as MQTT), receives an execution confirmation signal and updates the system power balance status.

[0051] For the second-level deviation, the edge node uploads the abnormal event data to the cloud collaborative management platform, triggers global optimization, and obtains the real-time output prediction of the photovoltaic array, the state of charge (SOC) of the energy storage, and the operating parameters of the power conversion system (PCS); constructs a mixed-integer programming model with the goal of minimizing the operating cost and maximizing the consumption of renewable energy, and the decision-making content includes the charge and discharge plan of the energy storage, the power curtailment ratio of the photovoltaic, and the switching time sequence of the adjustable load; decomposes the optimization result into control instructions for each node (such as the set value of the charge and discharge power of the energy storage and the output curve of the photovoltaic inverter), encrypts it and sends it to the edge execution terminal.

[0052] When the third-level deviation is detected, combined with the voltage / frequency monitoring data and topology analysis, locate the core load node causing the deviation, generate a load shedding sequence according to the load importance level (such as production line > lighting > charging pile), and preferentially cut off the low-priority load; send a tripping instruction to the circuit breaker controller and simultaneously start the standby power supply (such as a diesel generator) to supply power to the critical load. During this process, record the full-cycle data of the event, associate the device status log with the environmental mutation record, and generate a fault diagnosis report for operation and maintenance analysis.

[0053] Through the above technical solution, by parsing the load prediction value and calculating the deviation rate between it and the real-time load monitoring value, the load change situation can be accurately grasped. Based on the deviation rate, scheduling instructions are generated, and different scheduling strategies can be targeted according to the difference degree between the actual load and the predicted load, avoiding the blindness of scheduling, making the scheduling decision more scientific and accurate, and improving the stability and reliability of the microgrid operation; the local load transfer and cloud-side photovoltaic energy storage collaborative optimization in the hierarchical scheduling instructions can make full use of local resources and the coordination ability of the cloud side. The local load transfer can reasonably adjust the load distribution within the park microgrid to avoid local overload or light load; the cloud-side photovoltaic energy storage collaborative optimization can uniformly allocate resources such as photovoltaic arrays and energy storage converters from a global perspective to achieve the optimal utilization of photovoltaic energy storage resources, improve the energy utilization efficiency, and reduce the operation cost; in addition, by performing the protective load shedding operation, some non-critical loads can be cut off in time to ensure the safe and stable operation of the microgrid system, avoid system failures or power outages caused by problems such as overload, ensure the power supply reliability of important loads, enable the microgrid system to flexibly adopt corresponding scheduling measures according to different load deviation situations, and enhance the adaptability of the microgrid to external environmental changes; by combining the load prediction results with the real-time monitoring data, a closed-loop control from load prediction to scheduling instruction generation is realized, which can achieve the automatic monitoring, analysis and regulation of the microgrid, reduce manual intervention, and improve the management efficiency.

[0054] After generating the photovoltaic energy storage collaborative scheduling instruction according to the load prediction result, the method further includes: aggregating the model gradient parameters of multiple edge computing nodes on the cloud platform, and updating the global model parameters through a differential privacy protection mechanism and transmitting them back to the edge computing node to achieve the federated learning optimization of cloud-edge collaboration.

[0055] In an embodiment of the present specification, after completing the generation of the photovoltaic energy storage collaborative scheduling instruction, the load prediction model is continuously optimized through a cloud-edge collaborative federated learning mechanism to ensure that the model adapts to the dynamically changing microgrid environment under the premise of protecting data privacy. After locally training the spatio-temporal graph convolutional network at the edge computing node, the model gradient parameters (such as convolutional kernel weights, attention layer coefficients) are extracted. To protect data privacy, the following processing is performed on the original gradient: restricting the maximum norm of a single gradient to avoid the risk of privacy leakage caused by excessive gradient values; according to the preset differential privacy budget (ε value), calculating the amount of noise that satisfies (ε,δ)-differential privacy, and adding random noise that satisfies the Laplace distribution to the gradient parameters; performing homomorphic encryption processing on the noisy gradient to ensure that the original parameters cannot be parsed even if intercepted during the transmission process. The processed gradient is temporarily stored in the secure buffer area of the edge node and waits for the upload instruction.

[0056] After the cloud platform receives the encrypted gradients uploaded by multiple edge nodes, it performs secure aggregation, uses the Key Management Service (KMS) to decrypt the encrypted gradients uniformly, and aligns the gradient dimensions of each node according to the model layer structure. Calculate the cosine similarity matrix of all node gradients, and determine the gradient vectors with similarity lower than the threshold as outliers and eliminate them. Perform momentum weighted averaging on the valid gradients, and assign higher weights to the edge nodes with stable historical performance. The aggregation formula is: G t = βG t-1 +(1 - β)∑w i g i , where G t is the current global gradient, β is the momentum factor, w i is the node weight, and g i is the noise gradient; Apply a feasible region projection to the aggregated gradient to ensure that the update direction meets the model convergence constraint conditions. Update the cloud global model parameters based on the aggregated global gradient to generate a new version of the spatio-temporal graph convolutional network. The updated model parameters are transmitted back to the edge nodes through the following process. First, compare the difference between the old and new model parameters, and use sparse coding technology to compress the incremental parameters to reduce the amount of transmitted data. Then add a digital signature to the compressed incremental parameters to prevent tampering during transmission. Dynamically adjust the order and rate of parameter sharding according to the computing load and network status of the edge nodes, and give priority to ensuring the timely update of key nodes (such as areas with high photovoltaic penetration and large load fluctuations).

[0057] After the edge node receives the incremental parameters, it performs local model hot update, verifies the legality of the digital signature, and rejects illegal parameter injection. Decompress the incremental parameters and merge them with the local model parameters to generate an updated model. Retain the previous version of the model parameters. If the prediction error rate of the new model on the validation set exceeds the threshold, it will automatically roll back to the old version. And count metrics such as the prediction accuracy improvement rate and privacy loss (ε consumption) after the local model update, and feedback them to the cloud for optimizing the next round of federated learning strategy. The cloud platform dynamically adjusts the federated learning participation frequency of the edge nodes according to their historical contribution degree, calculates the contribution weight based on metrics such as the similarity between the node gradient and the global gradient, and the local prediction error improvement rate. Nodes with high contributions obtain more frequent participation qualifications; For nodes that are marked as outliers for multiple consecutive rounds, reduce their participation frequency until they are temporarily frozen.

[0058] Through the above technical solution, the cloud platform aggregates the model gradient parameters of multiple edge computing nodes, can integrate the data information of each node, and enables the update of global model parameters to fully consider the characteristics and conditions of different edge nodes. Thus, a more comprehensive and representative model can be obtained, which helps the model better capture complex patterns and rules in the data, improves the accuracy and generalization ability of the model, and further enhances the overall prediction and scheduling performance. Through the interaction of model parameters between the cloud and the edge, the cloud is responsible for aggregating and updating global model parameters, and the edge computing nodes provide local data and model gradient information, forming an efficient collaborative mechanism. The cloud can transmit the optimized global model parameters back to the edge computing nodes, enabling the edge nodes to use a better model for tasks such as load forecasting and scheduling instruction generation. At the same time, the edge nodes also provide data support for the update of the cloud model, realizing the complementary advantages of cloud-edge resources and improving the operation efficiency and intelligent level of the entire system.

[0059] Through the technical solution of the embodiments of this specification, traditional load forecasting methods mainly rely on the statistical characteristics of historical load curves. In the embodiments of this specification, by collecting multi-modal microgrid data, various data complement each other, enabling the model to more comprehensively consider various factors affecting the load, thereby effectively reducing the prediction error, improving the accuracy of load forecasting, and avoiding the problem that the prediction deviation of traditional methods grows exponentially under complex working conditions. Aiming at the problem that the spatial feature extraction in existing deep learning solutions relies on graph neural networks with fixed topologies and cannot adapt to the dynamic changes of topologies, in the embodiments of this specification, after collecting multi-modal microgrid data, abnormal events such as device commissioning and decommissioning can be monitored in real time through device status data. When the topology structure changes, it can trigger event-driven edge weight updates and locally update the adjacency matrix, thereby constructing a dynamic spatio-temporal correlation matrix, enabling the spatio-temporal graph convolutional network based on this matrix to adapt to the dynamic changes of the topology structure without retraining the model, and improving the adaptability of the model to the dynamic changes of the campus microgrid. In addition, the cloud centralized computing architecture leads to response delays exceeding real-time control requirements, and it is difficult for edge computing nodes to carry complex model deployments due to limited memory resources. In the embodiments of this specification, by collecting multi-modal microgrid data and processing it at the edge computing nodes, on the one hand, the time and communication cost of data transmission to the cloud can be reduced, and the response delay can be decreased; on the other hand, the collection and processing of multi-modal data can complete some operations such as feature extraction at the edge computing nodes, reducing the pressure on the memory resources of edge devices, enabling the edge computing nodes to better carry model operations, and meeting the requirements of real-time scheduling. At the same time, the load forecasting results based on multi-modal data can more accurately generate coordinated scheduling instructions for photovoltaic energy storage, such as generating hierarchical scheduling instructions according to the deviation rate between the load forecasting value and the real-time load monitoring value, realizing local load transfer, cloud-edge photovoltaic energy storage collaborative optimization, and protective load shedding operations, etc., improving energy utilization efficiency and the stability of system operation.

[0060] The embodiments of this specification also provide a park load prediction device based on multimodal data, such as Figure 2 shown, the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above method.

[0061] The embodiments of this specification also provide a non-volatile computer storage medium storing computer-executable instructions, and the computer-executable instructions are set to: execute the above method.

[0062] The various embodiments in this specification are all described in a progressive manner. For the same or similar parts among the various embodiments, reference can be made to each other. The key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the embodiments of the device, equipment, and non-volatile computer storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiments.

[0063] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in a different order than in the embodiments and still achieve the desired results. Additionally, the processes depicted in the figures do not necessarily require the particular order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0064] The device and medium provided by the embodiments of this specification correspond one-to-one with the method. Therefore, the device and medium also have beneficial technical effects similar to those of their corresponding methods. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the device and medium will not be elaborated here.

[0065] Those skilled in the art should understand that the embodiments of this specification can be provided as a method, a system, or a computer program product. Therefore, this specification can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] This specification is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the specification. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to produce a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for implementing the functions specified in one Figure 1 flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0067] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufacture including instruction means that implement the functions specified in one Figure 1 flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0068] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0069] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and a memory.

[0070] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer-readable media.

[0071] A computer-readable medium includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transitory medium that can be used to store information accessible by a computing device. As defined herein, a computer-readable medium does not include transitory computer-readable media such as modulated data signals and carrier waves.

[0072] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising the element.

[0073] The above description is only for one or more embodiments of this specification and is not intended to limit this specification. For those skilled in the art, there can be various changes and modifications to one or more embodiments of this specification. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of one or more embodiments of this specification shall be included within the scope of the claims of this specification.

Claims

1. A park load forecasting method based on multi-modal data, characterized in that, The method includes: Collecting multi-modal microgrid data through a multi-source heterogeneous sensor network deployed at a distribution node, where the multi-modal microgrid data includes electrical measurement data, environmental monitoring data, and equipment status data; Constructing a dynamic spatio-temporal correlation matrix based on the multi-modal microgrid data, and performing feature extraction based on the dynamic spatio-temporal correlation matrix through a spatio-temporal graph convolutional network deployed at an edge computing node to determine a spatio-temporal feature vector; Inputting the spatio-temporal feature vector into a time series fusion prediction model to output a load prediction result, and generating a coordinated scheduling instruction for photovoltaic energy storage according to the load prediction result, where the load prediction result includes a load prediction value and a confidence interval.

2. The method for predicting the park load based on multimodal data according to claim 1, wherein Constructing a dynamic spatio-temporal correlation matrix based on the multi-modal microgrid data, specifically including: Taking distribution transformers, photovoltaic arrays, and energy storage converters in a campus microgrid as graph nodes, calculating the equivalent impedance parameters between nodes according to the electrical measurement data in the multi-modal microgrid data to generate an initial adjacency matrix; Monitoring abnormal events according to the equipment status data in the multi-modal microgrid data, where the abnormal events include equipment connection / disconnection events or load power mutation events; When an abnormal event is detected, triggering an event-driven edge weight update to update the local adjacency matrix of the initial adjacency matrix and construct the dynamic spatio-temporal correlation matrix.

3. The method for predicting park load based on multi-modal data according to claim 2, wherein Calculating the equivalent impedance parameters between nodes according to the electrical measurement data in the multi-modal microgrid data to generate an initial adjacency matrix, specifically including: Extracting the real and imaginary parts of the fundamental voltage component in the electrical measurement data, calculating the absolute value of the phase difference between nodes, and deriving the equivalent impedance between nodes according to Ohm's law; Linearly combining the absolute value of the phase difference between nodes and the equivalent impedance between nodes according to a preset weight coefficient to generate a normalized distance index; Marking valid edges according to the physical connection status between nodes to eliminate redundant associated edges between non-directly connected nodes; Converting the normalized distance index into an adjacency matrix edge weight through an exponential mapping function to generate an initial adjacency matrix, where the edge weight and the normalized distance index in the exponential mapping function are negatively correlated.

4. A method for predicting park load based on multi-modal data according to claim 2, characterized in that After updating the local adjacency matrix of the initial adjacency matrix, the method further includes: Recording the last update timestamp of the edge weight of each adjacency matrix edge and calculating the update time interval between the current time and the last update timestamp; When the update time interval of a specified adjacency matrix edge is greater than a preset non-update duration, counting the historical adjacency matrix update frequency corresponding to the specified adjacency matrix edge; Applying an exponential decay factor to the edge weight of the specified adjacency matrix edge based on the historical adjacency matrix update frequency to determine a decay edge weight; When the decay edge weight is lower than a preset failure threshold, performing a removal operation on the specified adjacency matrix edge to generate an updated adjacency matrix.

5. A method for predicting the park load based on multi-modal data according to claim 1, characterized in that Performing feature extraction based on the dynamic spatio-temporal correlation matrix through a spatio-temporal graph convolutional network deployed at an edge computing node to determine a spatio-temporal feature vector, specifically including: Perform polynomial approximation decomposition on the dynamic spatio-temporal correlation matrix to generate the basis function coefficients of the graph convolution kernel, and input them into the spatio-temporal graph convolution network; Determine a multi-dimensional feature vector based on electrical measurement data, environmental monitoring data, and equipment status data, where the multi-dimensional feature vector includes node voltage phase difference, environmental temperature change rate, and equipment vibration spectrum characteristics; Input the multi-dimensional feature vector into the cross-modal attention layer to fuse the time series change rate corresponding to the environmental monitoring data, and generate the spatio-temporal feature vector.

6. A method for predicting the load of a park based on multi-modal data according to claim 1, characterized in that, Input the spatio-temporal feature vector into the time series fusion prediction model, and output the load prediction result, specifically including: Input the spatio-temporal feature vector into the time series fusion prediction model, analyze the feature importance of different time granularities through the quantile attention layer, and generate a prediction interval; Set a physical constraint term in the time series fusion prediction model to verify whether the initial prediction value conforms to the node power conservation law; If the initial prediction value does not conform to the node power conservation law, substitute the initial prediction value into the power flow equation to calculate the node voltage deviation; According to the node voltage deviation, adjust the prediction value distribution of the associated load nodes so that the voltage of all nodes is within the safe operating range and the power error rate is lower than the preset threshold.

7. A method for predicting the park load based on multimodal data according to claim 1, characterized in that, Generate a coordinated energy storage and photovoltaic (PV) dispatch instruction according to the load prediction result, specifically including: Analyze the load prediction value, and calculate the deviation rate between the load prediction value and the real-time load monitoring value corresponding to the electrical measurement data; Generate a hierarchical dispatch instruction according to the deviation rate, where the hierarchical dispatch instruction includes local load transfer, cloud-based coordinated energy storage and PV optimization, and protective load shedding operations.

8. A method for predicting park load based on multi-modal data according to claim 1, characterized in that, After generating the coordinated energy storage and PV dispatch instruction according to the load prediction result, the method further includes: Aggregate the model gradient parameters of multiple edge computing nodes on the cloud platform, and update the global model parameters through a differential privacy protection mechanism and send them back to the edge computing nodes to achieve the federated learning optimization of cloud-edge collaboration.

9. A park load prediction device based on multi-modal data, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method according to any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to: execute the method according to any one of claims 1-8.

Images