Power supply risk prediction method and system based on multi-source data fusion

Abstract

This invention discloses a method and system for predicting power supply risks based on multi-source data fusion, belonging to the field of intelligent operation and maintenance technology for power systems. The method includes: acquiring multi-source raw data from prediction units and constructing a standard fusion feature vector; combining the standard fusion feature vector with a pre-trained spatiotemporal risk-constrained prediction model to output a probabilistic prediction result containing multi-step future photovoltaic output and bus load prediction values; analyzing the probabilistic prediction result to construct a comprehensive risk index, and outputting a kilometer-level gridded risk heat map as the risk prediction result based on the comprehensive risk index. This invention achieves the fusion and probabilistic prediction of multi-source heterogeneous data, improving the ability to identify power supply risks and the level of precision in early warning under extreme weather scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, specifically to a method and system for predicting power supply risks based on multi-source data fusion. Background Technology

[0002] With the large-scale integration of new energy sources into the grid, the penetration rate of distributed power sources such as photovoltaics in the distribution network continues to increase, significantly increasing the uncertainty of power supply. At the same time, extreme weather events such as continuous rain, sudden thunderstorms, extreme heat waves, and cold waves occur frequently, placing higher demands on photovoltaic output and load forecasting. Accurately predicting distributed photovoltaic output and bus load, and assessing supply risks based on this, has become a core issue in the intelligent operation and maintenance of power systems.

[0003] Existing technologies mainly employ source-load joint forecasting methods driven by general weather forecasts and shallow statistical modeling methods based on historical time series. The former uses general weather forecasts as exogenous variables, combined with conventional machine learning models to predict photovoltaic output and load; the latter uses ARIMA, linear regression, or decision trees to model historical load or output sequences, capable of describing daily and weekly cycles with relatively low computational overhead. However, methods driven by general weather forecasts are insufficient in characterizing small-scale urban weather events and sudden weather events, lack uncertainty quantification and risk assessment mechanisms, and have low reliability under complex or extreme weather conditions, easily leading to misjudgments and omissions. Shallow statistical modeling methods based on historical time series are limited by single-source data and linear assumptions, making it difficult to integrate multi-source meteorological information and spatial differences. They also have weak responsiveness to nonlinear coupling relationships and abrupt changes, and cannot output probabilistic forecasts and gridded risk levels.

[0004] Therefore, there is an urgent need for a supply guarantee risk prediction method that can integrate multi-source heterogeneous data, capture spatiotemporal coupling characteristics, quantify and predict uncertainties, and output gridded risk warnings. Summary of the Invention

[0005] This invention addresses the technical problems in existing technologies, such as insufficient multi-source heterogeneous data fusion capabilities, low prediction accuracy in small-scale urban and extreme weather scenarios, and lack of uncertainty quantification and gridded risk assessment mechanisms. It provides a power supply risk prediction method and system based on multi-source data fusion.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for predicting power supply risks based on multi-source data fusion, including: Obtain multi-source raw data for the prediction unit, and construct a standard fusion feature vector based on the multi-source raw data; Combining the standard fusion feature vector with the pre-trained spatiotemporal risk constraint prediction model, the output is a probabilistic prediction result containing the photovoltaic power output prediction value and the bus load prediction value for multiple future steps. Analyze the probabilistic prediction results, construct a comprehensive risk index, and output a kilometer-level gridded risk heat map as the risk prediction result based on the comprehensive risk index.

[0007] Secondly, this invention provides a power supply risk prediction system based on multi-source data fusion, comprising: The feature construction module is used to acquire multi-source raw data of the prediction unit and construct a standard fusion feature vector based on the multi-source raw data; The probabilistic prediction module is used to combine the standard fused feature vector with the pre-trained spatiotemporal risk-constrained prediction model to output probabilistic prediction results that include the photovoltaic power output prediction value and the bus load prediction value for multiple future steps. The risk analysis output module is used to analyze the probabilistic prediction results, construct a comprehensive risk index, and output a kilometer-level gridded risk heat map as the risk prediction result based on the comprehensive risk index.

[0008] The beneficial effects of this invention are: Compared to existing technologies, this invention first fuses multi-source heterogeneous data, including meteorological observations, remote sensing data, numerical weather prediction, photovoltaic SCADA, bus load, and geographic topology, into a standard fusion feature vector, thus overcoming the shortcomings of single data sources in characterizing spatial differences and abrupt changes. Secondly, it combines the spatiotemporal risk-constrained prediction model to output probabilistic prediction results, achieving a quantitative expression of prediction uncertainty and improving prediction reliability under extreme weather scenarios. Thirdly, it calculates the probability of exceeding the upper threshold of load and the probability of exceeding the lower threshold of photovoltaic load based on the predicted distribution, and constructs a comprehensive risk index based on the exceedance magnitude, forming a multi-dimensional fusion risk quantification indicator. Finally, it outputs a kilometer-level gridded risk heat map, achieving gridded and refined early warning of supply security risks. This invention solves the problems of insufficient multi-source data fusion capabilities, low prediction accuracy under extreme weather conditions, and lack of uncertainty quantification and gridded risk assessment in existing technologies. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the power supply risk prediction method based on multi-source data fusion provided by this invention; Figure 2 This is a schematic diagram of the power supply risk prediction system based on multi-source data fusion provided by the present invention.

[0010] In the attached diagram, the components represented by each number are as follows: Feature construction module 11, probabilistic prediction module 12, risk analysis output module 13. Detailed Implementation

[0011] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for predicting power supply risks based on multi-source data fusion, including: S10: Obtain the multi-source raw data of the prediction unit, and construct a standard fusion feature vector based on the multi-source raw data; First, multi-source raw data is acquired for the prediction units. A prediction unit is the smallest spatial granularity for power supply risk prediction, representing an independent prediction area spatially divided according to substation power supply range, photovoltaic power station distribution, or kilometer-level grid. Prediction units are typically deployed in the power supply areas of various substations or at the locations of photovoltaic power stations within a prefecture-level power grid. Each prediction unit corresponds to a set of photovoltaic output and bus load data, used to independently predict the photovoltaic power generation capacity and load demand of its area. Through the spatial division of prediction units, the macro-level power supply guarantee problem can be decomposed into multiple refined prediction tasks for local areas, thereby acquiring multi-source raw data corresponding to each unit.

[0012] Specifically, the multi-source raw data is a heterogeneous collection of information from different data sources, including at least meteorological observation data, remote sensing data, numerical weather prediction data, photovoltaic power generation operation data, bus load data, and power grid geographic topology data, which is used to provide a comprehensive data foundation for subsequent feature construction and model prediction.

[0013] Secondly, a standardized fusion feature vector is constructed based on multi-source raw data. Since the multi-source raw data originates from different data sources, each with different temporal resolution, spatial resolution, sampling frequency, and units of measurement—for example, meteorological observation data is at the hourly level, remote sensing data at the minute level, and numerical weather prediction at the three-hour level—and the numerical ranges of each data point vary significantly, directly concatenating them would make it difficult for the model to effectively learn the feature relationships across data sources. Therefore, it is necessary to perform spatiotemporal alignment, missing value imputation, outlier detection, and normalization on the multi-source raw data. This unifies data from different sources and at different scales into the same spatiotemporal grid and unit of measurement to construct a standardized fusion feature vector. This provides input features with a unified format, clear semantics, and consistent units of measurement for subsequent spatiotemporal risk constraint prediction models.

[0014] Specifically, the execution steps of S10 are as follows: acquiring multi-source raw data of the prediction unit, and constructing a standard fusion feature vector based on the multi-source raw data, including: The multi-source raw data is obtained through multiple data sources, wherein the multi-source raw data includes at least meteorological observation data, remote sensing data, numerical weather prediction data, photovoltaic power generation operation data, bus load data, and power grid geographic topology data. The multi-source raw data is subjected to spatiotemporal alignment, missing value imputation, outlier detection and normalization, and then concatenated into the standard fusion feature vector. The standard fusion feature vector includes meteorological element components, photovoltaic power output historical sequence components, load historical sequence components, static geographic equipment feature components, and calendar feature components.

[0015] First, multiple sources of raw data are acquired through multi-source data sources. These sources include at least meteorological observation data, remote sensing data, numerical weather prediction data, photovoltaic power generation operation data, bus load data, and power grid geographic topology data.

[0016] Specifically, meteorological observation data comes from ground-based meteorological stations and includes elements such as temperature, irradiance, wind speed, relative humidity, and precipitation intensity, used to characterize the actual meteorological conditions of the forecast unit. Remote sensing data comes from satellite cloud images and radar echoes, including information such as cloud cover, convection identification, and precipitation estimation, used to capture the spatial distribution of cloud cover and precipitation intensity. Numerical weather forecast data comes from global or regional numerical weather prediction models, including forecast values ​​of meteorological elements for multiple future hours, used to provide forward-looking meteorological input. Photovoltaic power generation operation data comes from the SCADA system of photovoltaic power plants, including historical active power output sequences and installed capacity, used to reflect the actual operating status of photovoltaic power output. Bus load data comes from grid bus SCADA or PMU measurements, including historical load sequences, used to reflect the electricity demand of each power supply area. Grid geographic topology data comes from grid GIS or CIM models, including substation locations, line connection relationships, and admittance matrices, used to construct spatial adjacency relationships.

[0017] Secondly, the obtained multi-source raw data undergoes spatiotemporal alignment, missing value imputation, outlier detection, and normalization, and is then concatenated into a standard fusion feature vector. Spatiotemporal alignment unifies data from different sources with varying temporal and spatial resolutions to the same temporal sampling interval and spatial grid system. For example, hourly meteorological observation data, minute-level satellite remote sensing data, and three-hour numerical weather prediction data are unified to a 15-minute time interval using linear interpolation or nearest-neighbor interpolation. Point data from meteorological stations are mapped to each prediction unit grid using inverse distance weighted interpolation. Missing value imputation fills in missing values ​​caused by sensor failures or transmission interruptions. For a small number of isolated missing points, linear interpolation of adjacent valid time points is used; for consecutive missing points, forward or backward imputation is used; and missing values ​​at both ends are marked as invalid and not included in subsequent statistics. Outlier detection is the process of identifying and handling extreme values ​​in data that deviate from the normal distribution range. It involves calculating the z-score by taking the mean and standard deviation of each feature dimension. When the absolute value of the z-score is greater than 3, it is considered an outlier. The value is then brought back to a threshold range of the mean plus or minus three times the standard deviation, or marked as missing and handled through imputation. Normalization is the process of scaling features with different dimensions and numerical ranges to the same scale range. The Min-Max normalization method is used, calculating the minimum and maximum values ​​of each continuous feature dimension on the training set. The original value is then subtracted from the minimum value and divided by the difference between the maximum and minimum values, ensuring that the normalized value falls within the range of 0 to 1, thus eliminating the impact of dimensional differences on model training.

[0018] After completing the above processing, all feature components are concatenated in a preset order to form a standard fused feature vector. This standard fused feature vector includes meteorological element components, photovoltaic power output historical sequence components, load historical sequence components, static geographic equipment feature components, and calendar feature components.

[0019] Specifically, the meteorological component includes meteorological variables such as temperature, irradiance, wind speed, relative humidity, and precipitation intensity, both at the current time and within the historical window. The photovoltaic output historical sequence component includes time-series segments of photovoltaic output of the forecast unit over a past period. The load historical sequence component includes time-series segments of bus load of the forecast unit over a past period. The static geographic equipment characteristic component includes static attributes of the forecast unit that do not change over time, such as installed capacity, wiring method, latitude and longitude, and land type. The calendar characteristic component includes time-related features such as hour, day of the week, and holiday markings.

[0020] The "past period" refers to the length of the time window used to construct the historical sequence segments. This time window length is determined comprehensively based on the diurnal periodicity of photovoltaic power output and load, as well as the model's ability to capture historical dependencies. For example, it can be set to 24 hours, corresponding to 96 time steps. When the time step is 15 minutes, the historical sequence of the past 24 hours contains 96 sampling points, which can cover the complete diurnal periodic variation pattern and provide the model with sufficient historical context information.

[0021] Through the above standardization process and feature concatenation, a standardized fusion feature vector with uniform format, consistent dimensions, and clear semantics can be obtained, providing high-quality input for subsequent spatiotemporal risk constraint prediction models.

[0022] S20: Combining the standard fusion feature vector with the pre-trained spatiotemporal risk constraint prediction model, outputting a probabilistic prediction result that includes the photovoltaic power output prediction value and the bus load prediction value for multiple future steps; Secondly, combining the aforementioned standard fusion feature vector with the pre-trained spatiotemporal risk-constrained prediction model, a probabilistic prediction result containing the predicted photovoltaic power output and bus load for multiple future time steps is output. The pre-trained spatiotemporal risk-constrained prediction model is an end-to-end prediction model built on a deep neural network. It extracts the meteorological spatiotemporal features, historical power characteristics, and geographical static features of the prediction unit based on the input standard fusion feature vector. Through spatiotemporal coding and a cross-modal attention mechanism, it fuses multimodal information to output the predicted photovoltaic power output and bus load for multiple future time steps, and characterizes the prediction uncertainty using a Gaussian distribution parameterization.

[0023] Specifically, this spatiotemporal risk-constrained prediction model first captures temporal dependencies in historical sequences using a time encoder, and then fuses spatial correlation information between adjacent prediction units using a spatial feature aggregator to obtain a comprehensive spatiotemporal representation vector. Subsequently, a three-branch decoder outputs probabilistic prediction results for meteorological elements, photovoltaic power output, and bus load, respectively. The output layers of the photovoltaic power output and load prediction branches are configured with a mean and standard deviation dual-head structure, outputting the predicted mean and standard deviation, respectively, forming a Gaussian distribution for probabilistic predictions. Through this spatiotemporal risk-constrained prediction model, point prediction values ​​and prediction confidence intervals for future multi-step photovoltaic power output and bus load can be obtained simultaneously, providing complete probabilistic distribution information for subsequent risk quantification.

[0024] Specifically, by combining the standard fused feature vector with the pre-trained spatiotemporal risk-constrained prediction model, a probabilistic prediction result is output, containing predictions of photovoltaic power output and bus load for multiple future steps, including: By using a priori temporal encoder and spatial feature aggregator, the standard fused feature vector is spatiotemporally encoded to obtain a comprehensive spatiotemporal representation vector; The spatiotemporal risk-constrained prediction model is used to perform probabilistic prediction of the comprehensive spatiotemporal representation vector based on a three-branch decoding channel, wherein the three-branch decoding channel includes a meteorological prediction branch, a photovoltaic power output prediction branch, and a load prediction branch. The outputs of the three-branch decoding channels are subjected to inter-modal adaptive fusion to obtain probabilistic prediction results of photovoltaic power output prediction and bus load prediction. The probabilistic prediction results are output in Gaussian distribution parameterized form, including the predicted mean and predicted standard deviation of the photovoltaic power output prediction value and the bus load prediction value.

[0025] First, a spatiotemporal encoding of the standard fused feature vector is performed using a prior temporal encoder and a spatial feature aggregator to obtain a comprehensive spatiotemporal representation vector. The temporal encoder, used to capture temporal dependencies in historical sequences, can be implemented using a bidirectional long short-term memory network, a temporal convolutional network, or a temporal Transformer.

[0026] For example, taking a bidirectional long short-term memory (LSTM) network as an example, the specific architecture of the time encoder is as follows: A two-layer bidirectional LSM network is set up, with 128 hidden units in each layer. The input layer receives the historical sequence component from the standard fused feature vector. The time step of this sequence is 96, corresponding to sampling points at 15-minute intervals over the past 24 hours. The feature dimension of each time step is the input feature dimension d. The first-layer bidirectional LSM network receives the input sequence. The forward LSM network calculates the hidden state sequentially from step 1 to step 96, and the reverse LSM network calculates the hidden state sequentially from step 96 to step 1. The output of each time step is the concatenation of the forward and reverse hidden states, with a dimension of 256. The second-layer bidirectional LSM network receives the output of the first layer and is also calculated bidirectionally, with an output dimension of 256. The concatenated hidden state of the last time step of the second layer is embedded as a time feature, with a dimension of 256. The time encoder enables the model to utilize contextual information from both historical and future moments, effectively extracting daily cyclical patterns and trend changes in photovoltaic power output and load sequences.

[0027] In addition, spatial feature aggregators are used to fuse spatial correlation information between adjacent prediction units, and can be implemented using graph convolutional networks or graph attention networks.

[0028] For example, taking a graph attention network as an example, the specific architecture of the spatial feature aggregator is as follows: Two graph attention layers are set, each with four attention heads. First, an adjacency matrix is ​​constructed based on the power grid geographic topology data. Using prediction units as nodes, edges are determined according to the line connection relationships between substations. If there is a direct line connection between two prediction units, the corresponding position in the adjacency matrix is ​​set to 1; otherwise, it is set to 0. Simultaneously, edge weights can be assigned based on the reciprocal of the electrical distance or geographical distance. The input feature is the temporal feature embedding output by the time encoder, with a dimension of 256, corresponding to U prediction units. The first graph attention layer calculates the attention coefficient between each node and its neighboring nodes. For node i and its neighbor node j, the attention coefficient is obtained by concatenating the feature vectors of node i and node j, mapping them through a single-layer feedforward network to obtain a scalar, and then normalizing it using Softmax to obtain the attention weight. Each attention head independently calculates a set of attention weights and performs a weighted sum of the features of its neighboring nodes to obtain the output of that head. The outputs of the four attention heads are concatenated to obtain the output of the first graph attention layer, with a dimension of 256 multiplied by 4 equaling 1024. The second graph attention layer has the same structure, with an input dimension of 1024 and an output dimension of 256. The output of the second graph attention layer is embedded as a spatial feature with a dimension of 256. Through the graph attention network, the model can adaptively learn the importance of different neighboring nodes in predicting the target node, effectively fusing spatial correlation information between adjacent prediction units, making it particularly suitable for areas with sparse meteorological stations or unevenly distributed photovoltaic power plants.

[0029] The temporal feature embedding (dimension 256), spatial feature embedding (dimension 256), and the original standard fused feature vector (dimension d) are concatenated to obtain a comprehensive spatiotemporal representation vector with a dimension of 512 plus d. This comprehensive spatiotemporal representation simultaneously contains historical dynamic information, spatial correlation information, and original feature information, providing a rich feature foundation for subsequent multi-branch decoding.

[0030] Specifically, the standard fused feature vector is spatiotemporally encoded using a priori temporal encoder and spatial feature aggregator to obtain a comprehensive spatiotemporal representation vector, including: A temporal encoder is used to extract temporal features from the standard fused feature vector to obtain temporal feature embeddings. Using the adjacency matrix constructed based on power grid geographic topology data as a constraint, the spatial feature embedding is obtained by aggregating spatial dimension features of the temporal feature embedding through the spatial feature aggregator; The temporal feature embedding, the spatial feature embedding, and the standard fused feature vector are concatenated to obtain a comprehensive spatiotemporal representation; The temporal encoder is implemented based on any one of bidirectional LSTM, temporal convolutional network, and temporal Transformer, and the spatial feature aggregator is implemented based on any one of graph convolution and graph attention.

[0031] First, a time-series encoder is used to extract time-dimensional features from the standard fused feature vector, resulting in a time-series feature embedding. The standard fused feature vector includes meteorological element components, historical photovoltaic output sequence components, historical load sequence components, static geographic equipment feature components, and calendar feature components, among which the historical sequence components exhibit a clear temporal order. The time-series encoder is used to capture long-term dependencies and short-term fluctuation patterns within the sequence components.

[0032] Specifically, taking a bidirectional long short-term memory network as an example, the time encoder receives historical sequence segments arranged by time steps from the standard fused feature vector. It processes the sequence data sequentially from front to back through a forward recurrent unit, extracting the hidden state at each time step. Simultaneously, it processes the sequence data from back to front through a backward recurrent unit, extracting the reverse hidden state. The forward and reverse hidden states are concatenated at each time step to obtain a time feature embedding containing past and future contextual information. This time feature embedding can effectively characterize the daily cycle patterns, trend changes, and sudden fluctuations in the historical sequences of photovoltaic power output and load.

[0033] Secondly, using the adjacency matrix constructed based on power grid geographic topology data as a constraint, a spatial feature aggregator is used to aggregate the spatial dimension features of the temporal feature embeddings, resulting in spatial feature embeddings. Specifically, the power grid geographic topology data includes information such as substation locations, line connection relationships, and admittance matrices, from which an adjacency matrix can be constructed. The adjacency matrix is ​​a square matrix, with its rows and columns corresponding to each prediction unit. Matrix elements represent the spatial correlation strength between two corresponding prediction units; for example, non-zero element values ​​can be determined based on the reciprocal of geographical distance, electrical distance, or topological connection relationships. The spatial feature aggregator, constrained by the adjacency matrix, performs spatial aggregation on the temporal feature embeddings of each prediction unit, weighting and fusing the features of the unit itself and its neighboring units.

[0034] Specifically, taking graph attention networks as an example, the spatial feature aggregator first calculates the attention weight between each prediction unit and each of its neighboring units. This weight depends on the similarity or correlation between the temporal feature embeddings of the two units. Then, the temporal feature embeddings of the neighboring units are weighted and summed according to the attention weights to obtain the spatial feature embedding of the unit. Through spatial feature aggregation, information between adjacent prediction units can be shared, enabling the model to use observation data from the spatial neighborhood to correct the prediction of the target unit, which is particularly suitable for areas with sparse meteorological stations or uneven distribution of photovoltaic power plants.

[0035] Finally, the temporal feature embedding, spatial feature embedding, and standard fused feature vector are concatenated to obtain a comprehensive spatiotemporal representation. Specifically, the temporal feature embedding reflects the dynamic patterns of the historical sequence, the spatial feature embedding reflects the spatial correlation information of adjacent units, and the standard fused feature vector retains the original static features and calendar features. The concatenation of these three forms a higher-dimensional comprehensive spatiotemporal representation vector, which simultaneously contains temporal dynamic information, spatial correlation information, and original input information, providing a rich feature foundation for the subsequent three-branch decoding channels.

[0036] Through the aforementioned spatiotemporal coding, the spatiotemporal risk-constrained prediction model can simultaneously capture the periodic variation patterns of photovoltaic power output and load in the time dimension, as well as the inter-regional coupling relationships in the spatial dimension, thereby improving the prediction accuracy under complex meteorological scenarios.

[0037] The temporal encoder is implemented based on any one of bidirectional LSTM, temporal convolutional networks, or temporal Transformers, while the spatial feature aggregator is implemented based on any one of graph convolutional networks or graph attention networks. Different implementation methods can be chosen based on the actual data scale and computational resources. For example, for long sequence prediction tasks, temporal Transformers can be used to better capture global dependencies, while for large-scale spatial graph structures, graph attention networks can be used to adaptively learn the weight distribution between nodes.

[0038] Furthermore, a spatiotemporal risk-constrained prediction model is used to perform probabilistic prediction of the comprehensive spatiotemporal representation vector based on a three-branch decoding channel. This three-branch decoding channel includes a meteorological prediction branch, a photovoltaic power output prediction branch, and a load prediction branch.

[0039] Specifically, the training and construction process of the spatiotemporal risk constraint prediction model includes: Acquire training sample data, wherein the training sample data includes a sample standard fusion feature vector and a corresponding supervision target vector, and the supervision target vector includes the true values ​​of meteorological elements, photovoltaic power output, and bus load; The three-branch decoding channel is constructed based on a deterministic neural network, wherein the decoding output layer of each branch is set as a distributed parameterized output head, the distributed parameterized output head includes a mean output subheader and a standard deviation output subheader, and the standard deviation output subheader is constrained to have a positive output value by an activation function; The time encoder, the spatial feature aggregator, the three-branch decoding channel, and the cross-modal attention mechanism are integrated to form the overall network structure of the spatiotemporal risk constraint prediction model. The spatiotemporal risk constraint prediction model is trained in a supervised manner using the training sample data and a hybrid loss function. In this cross-modal attention mechanism, the query vector comes from the feature representation of the meteorological forecasting branch, and the key vector and value vector come from the feature representations of the photovoltaic power output forecasting branch and the load forecasting branch. The adaptive allocation of intermodal weights is achieved through scaling dot product attention calculation.

[0040] First, training sample data is acquired. This data includes a standard fusion feature vector and a corresponding supervisory target vector. The supervisory target vector contains the true values ​​of meteorological elements, photovoltaic power output, and bus load. Specifically, this training sample data originates from multi-source raw data within a historical time period. After spatiotemporal alignment, cleaning, normalization, and feature concatenation in step S10, a standard fusion feature vector is formed. The supervisory target vector represents the future true value corresponding to the input feature vector in time. For example, for a standard fusion feature vector input at time t, the supervisory target is the actual observed values ​​of meteorological elements, photovoltaic power output, and load in the next τ steps. The next τ steps refer to the number of future time steps to be predicted. Each time step corresponds to a fixed time interval. This prediction step size is set comprehensively based on the operational response time requirements of power supply risk warning and the fluctuation characteristics of photovoltaic power output and load. For example, it can be set to 96 steps. When the time step size is 15 minutes, 96 steps correspond to a prediction window of the next 24 hours, which can cover the complete daily cycle variation pattern and meet the operational requirements for advance warning in power supply risk warning. The training samples are divided into training set, validation set and test set in chronological order, for example, in a ratio of 7:2:1.

[0041] Secondly, a three-branch decoding channel is constructed based on a deterministic neural network, where the decoding output layer of each branch is set as a distributed parameterized output head. This distributed parameterized output head includes a mean output subheader and a standard deviation output subheader. The standard deviation output subheader is constrained by an activation function to output a positive value. The mean output subheader uses a linear activation function to directly output the predicted mean, which represents the model's most likely estimate of future photovoltaic power output or load. The standard deviation output subheader uses a Softplus activation function, where the standard deviation equals the natural logarithm of one plus e to the power of x, ensuring a positive output value. This standard deviation represents the model's degree of uncertainty regarding the prediction. Through the mean and standard deviation, the model can output probabilistic prediction results in the form of a Gaussian distribution, where the mean represents the best estimate of the prediction, and the standard deviation represents the dispersion or confidence range of the prediction.

[0042] Specifically, the three-branch decoding channel includes a weather forecast branch, a photovoltaic power output forecast branch, and a load forecast branch. The three branches share the comprehensive spatiotemporal representation vector extracted by the encoder and fine-tune their parameters for their respective tasks. The decoder for each branch can be implemented using a multi-layer fully connected network. For example, each branch has two fully connected layers: the first layer contains 256 neurons, and the second layer contains 128 neurons. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function. The output layer splits into two sub-heads: a mean head outputting the predicted mean, and a standard deviation head outputting the predicted standard deviation. The mean head has the same number of neurons as the target dimension and uses a linear activation function; the standard deviation head has the same number of neurons as the target dimension and uses a Softplus activation function to constrain the output to be positive. Through these distributed parameterized output heads, the model can output probabilistic prediction results, providing complete distribution information for subsequent risk quantification, including predicted values ​​for each quantile and prediction confidence intervals.

[0043] Furthermore, an integrated temporal encoder, spatial feature aggregator, three-branch decoding channel, and cross-modal attention mechanism constitute the overall network structure of the spatiotemporal risk-constrained prediction model. The temporal encoder and spatial feature aggregator form the encoder section, responsible for extracting a comprehensive spatiotemporal representation from the input standard fused feature vector. The three-branch decoding channel forms the decoder section, responsible for decoding probabilistic predictions of meteorology, photovoltaic power generation, and load from the comprehensive spatiotemporal representation. The cross-modal attention mechanism is embedded between the branches of the decoder to achieve interactive fusion of information between modalities. Specifically, in the middle layer of the decoder, the feature representation of the meteorological prediction branch undergoes a linear transformation to obtain a query vector, while the feature representations of the photovoltaic output prediction branch and the load prediction branch undergo linear transformations to obtain key and value vectors. Through scaled dot product attention calculation, the meteorological branch can adaptively adjust the weights of its features according to the features of the photovoltaic and load branches, achieving adaptive weight allocation between modalities.

[0044] Furthermore, by combining the training sample data, the spatiotemporal risk constraint prediction model is trained in a supervised manner using a hybrid loss function.

[0045] The hybrid loss function is a weighted combination of three terms: mean squared error, quantile loss, and risk penalty, including: The mean square error between the predicted mean and the true value is defined as the mean square error term. Based on the preset list of quantiles, the predicted value of each quantile is calculated by combining the predicted mean and the predicted standard deviation, and the average of the loss between the predicted quantile value and the true quantile value is defined as the quantile loss term. The risk penalty term is defined by combining a preset relative error threshold and an activation function. The risk penalty term is used to generate a penalty gradient through the activation function when the prediction deviation exceeds the relative error threshold.

[0046] Specifically, the hybrid loss function comprises a weighted combination of three terms: mean squared error, quantile loss, and risk penalty. The mean squared error term constrains the closeness of the predicted mean to the true value, ensuring basic prediction accuracy. The quantile loss term drives the model to output unbiased estimates of each quantile, contributing to the construction of the cumulative distribution function of the predictions. The risk penalty term generates a penalty gradient when the prediction deviation exceeds a set threshold, guiding the model to provide more conservative predictions in high-risk regions.

[0047] First, the mean squared error between the predicted mean and the true value is defined as the mean squared error term. The mean squared error term calculates the average of the squares of the differences between the predicted mean and the true value, used to constrain the model's output predicted mean to be as close as possible to the true value, ensuring basic prediction accuracy. The mean squared error term applies a squared penalty to larger deviations, guiding the model to prioritize reducing significant errors.

[0048] Secondly, based on the preset quantile list, the predicted quantile value for each quantile is calculated using the predicted mean and predicted standard deviation. The quantile loss term is defined as the average loss between the predicted and true quantile values. The preset quantile list can be set according to business needs, for example, it can be set to five quantiles: 0.1, 0.25, 0.5, 0.75, and 0.9. For each quantile q, the predicted quantile value equals the predicted mean plus the predicted standard deviation multiplied by the z-score corresponding to quantile q under the standard normal distribution. Quantile loss, also known as quantile percentage loss, is based on the core idea of ​​applying asymmetric penalty weights to over-predictions and under-predictions. Specifically, let e be the deviation between the predicted and true values. When the predicted value is higher than the true value, the loss is (1-q) multiplied by the deviation e; when the predicted value is lower than the true value, the loss is q multiplied by the absolute value of the deviation e. Taking q = 0.9 as an example, this quantile imposes a high-weight penalty of 0.9 on underestimating predictions and a low-weight penalty of only 0.1 on overestimating predictions. Therefore, the model tends to output overestimating predictions, used to characterize the upper limit risk. Taking q = 0.1 as another example, this quantile imposes a high-weight penalty of 0.9 on overestimating predictions and a low-weight penalty of only 0.1 on underestimating predictions. Therefore, the model tends to output underestimating predictions, used to characterize the lower limit risk. Through quantile loss, the model can learn the conditional quantiles of the prediction results, thereby constructing a complete cumulative distribution function, providing a distributional basis for subsequent over-limit probability calculations.

[0049] Furthermore, combining a preset relative error threshold and activation function, a risk penalty term is defined. This risk penalty term is used to generate a penalty gradient through the activation function when the prediction deviation exceeds the relative error threshold. The relative error threshold can be set according to the business requirements for prediction accuracy, for example, it can be set to 20% of the prediction mean or the 90th percentile of the training set error. When the absolute value of the prediction deviation does not exceed the threshold, the risk penalty term is zero; when the absolute value of the prediction deviation exceeds the threshold, the activation function generates a positive penalty gradient, and the larger the deviation, the stronger the penalty.

[0050] The commonly used activation function is the linear rectified function, which means that the output equals the input when the input is greater than zero, and the output is zero when the input is less than or equal to zero. Through a risk penalty term, the model primarily optimizes mean squared error and quantile loss within the normal prediction range. In extreme deviation scenarios, an additional penalty is applied, guiding the model to provide more conservative and accurate predictions in high-risk areas, thus avoiding omissions in supply guarantee warnings due to underestimating risks.

[0051] Finally, the mean squared error term, quantile loss term, and risk penalty term are weighted and summed according to preset weights to obtain the total loss value of the hybrid loss function. For example, the weight of the mean squared error term can be set to 1.0, the weight of the quantile loss term can be set to 0.5, and the weight of the risk penalty term can be set to 0.3. Through this hybrid loss function, the model can output a calibrated probability distribution and remain sensitive to extreme deviations while maintaining basic prediction accuracy, thus improving prediction reliability and risk identification capabilities in complex weather scenarios.

[0052] Specifically, during training, the standard fused feature vectors of the training samples are input into the model to obtain the predicted mean and standard deviation of each branch. The mixed loss value is calculated, and the network weight parameters are iteratively updated using the backpropagation algorithm and optimizer. After each training epoch, the model performance is evaluated using a validation set. Training stops when the validation set loss no longer decreases for several consecutive epochs. For example, if the validation set loss value does not fall below the previously recorded best loss value for 10 consecutive epochs, the model is considered to have converged, training stops, and the model parameters corresponding to the minimum validation set loss are saved, resulting in the trained spatiotemporal risk constraint prediction model. This early stopping strategy effectively avoids model overfitting while ensuring that the model achieves optimal generalization performance on the validation set.

[0053] In this cross-modal attention mechanism, the query vector comes from the feature representation of the meteorological forecasting branch, while the key and value vectors come from the feature representations of the photovoltaic output forecasting branch and the load forecasting branch. Adaptive weight allocation between modalities is achieved through scaled dot product attention calculation. Specifically, the calculation formula for the cross-modal attention mechanism is: the attention output equals the attention weight matrix multiplied by the value vector, where the attention weight matrix is ​​obtained by scaling the dot product of the query vector and the key vector and then normalizing it using Softmax. This mechanism enables the meteorological branch to dynamically adjust the importance of its features based on the current state of the photovoltaic and load branches. For example, when a sudden drop in irradiance leads to a sharp decrease in photovoltaic output, the feature changes in the photovoltaic branch are transmitted to the meteorological branch through the attention mechanism, causing the meteorological branch to correspondingly increase its weight in irradiance prediction, thereby correcting the predictions for subsequent time steps.

[0054] Specifically, a spatiotemporal risk-constrained prediction model is used to perform probabilistic predictions on the comprehensive spatiotemporal representation vector based on a three-branch decoding channel. The three-branch decoding channel includes a meteorological prediction branch, a photovoltaic output prediction branch, and a load prediction branch. These three branches share the comprehensive spatiotemporal representation vector extracted by the encoder and fine-tune their parameters for their respective tasks. The meteorological prediction branch decodes and reconstructs the comprehensive spatiotemporal representation vector, outputting probabilistic predictions of meteorological elements for multiple future steps, including the predicted mean and standard deviation of meteorological variables such as temperature, irradiance, and wind speed. The photovoltaic output prediction branch receives the comprehensive spatiotemporal representation vector and decodes and reconstructs it in conjunction with the output features of the meteorological prediction branch, outputting probabilistic predictions of photovoltaic output for multiple future steps. The load prediction branch receives the comprehensive spatiotemporal representation vector, decodes and reconstructs it, and outputs probabilistic predictions of bus load for multiple future steps. The output layer of each branch is configured as a dual-head structure of mean and standard deviation. The mean head outputs the predicted mean, and the standard deviation head outputs a positive value after being constrained by an activation function, together forming a probabilistic prediction in Gaussian distribution form, enabling the model to quantitatively express the uncertainty of the prediction.

[0055] Specifically, the spatiotemporal risk constraint prediction model is used to perform probabilistic prediction of the integrated spatiotemporal representation vector based on a three-branch decoding channel, including: The meteorological forecasting branch receives the integrated spatiotemporal representation vector, decodes and reconstructs it, and outputs the probabilistic forecast results of meteorological elements in the form of Gaussian distribution parameterization for the next multiple steps. The load forecasting branch receives the integrated spatiotemporal representation vector, decodes and reconstructs it, and outputs the probabilistic prediction result of the bus load in the parameterized form of Gaussian distribution for the next multiple steps. The photovoltaic output prediction branch receives the integrated spatiotemporal representation vector and the probabilistic prediction results of the meteorological elements, decodes and reconstructs them, and outputs the probabilistic prediction results of photovoltaic output in the form of Gaussian distribution parameterization for the next many steps.

[0056] First, the meteorological forecasting branch receives the integrated spatiotemporal representation vector, decodes and reconstructs it, and outputs probabilistic forecasts of meteorological elements in Gaussian distributed parameterized form for multiple future time steps. The integrated spatiotemporal representation vector contains temporal dynamics, spatial correlation, and original feature information. The meteorological forecasting branch decodes this vector through a multi-layer fully connected network, outputting meteorological element forecasts for multiple future time steps. Taking temperature as an example, the meteorological forecasting branch outputs the forecast mean and forecast standard deviation for each forecast time step, forming a Gaussian distribution. The forecast mean represents the most likely value of the meteorological element at that moment, and the forecast standard deviation represents the degree of uncertainty in the forecast; a larger standard deviation indicates more drastic changes in meteorological conditions or a lower confidence level in the model's prediction for that moment. The probabilistic forecasts of meteorological elements include the distribution parameters of variables such as temperature, irradiance, wind speed, relative humidity, and precipitation intensity, providing crucial exogenous inputs for the photovoltaic output forecasting branch.

[0057] Simultaneously, the load forecasting branch receives the integrated spatiotemporal representation vector, decodes and reconstructs it, and outputs the probabilistic bus load forecast results in Gaussian distribution parameterized form for multiple future time steps. The load forecasting branch independently decodes the integrated spatiotemporal representation vector, outputting the mean and standard deviation of the bus load forecast for multiple future time steps. Load forecasting is mainly affected by historical load patterns, temperature, holidays, and other factors, which are already included in the integrated spatiotemporal representation vector. By outputting load forecasts in Gaussian distribution form, the model can quantify the uncertainty of load forecasting. For example, when air conditioning load surges under extreme high temperatures, the forecast standard deviation will increase accordingly, reflecting the model's uncertain estimation of the timing and magnitude of load peaks.

[0058] Furthermore, the photovoltaic (PV) output prediction branch receives the integrated spatiotemporal representation vector and the probabilistic prediction results of meteorological elements, decodes and reconstructs them, and outputs the probabilistic prediction results of PV output in Gaussian distribution parameterized form for the next several steps. PV output is strongly coupled with meteorological conditions, especially irradiance; therefore, the PV output prediction branch not only receives the integrated spatiotemporal representation vector but also additionally receives the probabilistic prediction results of meteorological elements output by the meteorological prediction branch as conditional input. Specifically, the feature representation of the intermediate layer of the meteorological prediction branch, after linear transformation, is concatenated or weighted and fused with the integrated spatiotemporal representation vector before being input into the decoding network of the PV output prediction branch. By incorporating meteorological prediction results, the PV output prediction branch can anticipate trends in irradiance changes. For example, when the meteorological branch predicts a sharp drop in irradiance in the next two hours, the output mean of the PV output prediction branch will be adjusted downwards accordingly, and the prediction standard deviation will increase accordingly to reflect the uncertainty of output reduction due to cloud cover.

[0059] Through the collaborative prediction of the three-branch decoding channels, the model can achieve joint modeling of meteorology, photovoltaics and load. The branches can mutually verify and correct each other through cross-modal attention mechanism, thereby improving the prediction accuracy and robustness in complex meteorological scenarios.

[0060] Furthermore, the outputs of the three-branch decoding channels are subjected to inter-modal adaptive fusion to obtain probabilistic prediction results of photovoltaic power output prediction and bus load prediction. Since photovoltaic power output is strongly coupled with meteorological conditions and load is closely related to calendar factors such as temperature and holidays, there is a mutual verification and correction relationship between the prediction results of each branch. Therefore, a cross-modal attention mechanism is introduced to achieve adaptive allocation of inter-modal weights.

[0061] Specifically, the cross-modal attention mechanism is defined as follows: the attention output equals the attention weight matrix multiplied by the value vector, where the attention weight matrix is ​​obtained by scaling the dot product of the query vector and the key vector and then normalizing it using Softmax. The query vector comes from the feature representation of the meteorological forecast branch, while the key and value vectors come from the feature representations of the photovoltaic output forecast branch and the load forecast branch. Through this mechanism, when the irradiance prediction of the meteorological branch is too low, the output forecast weight of the photovoltaic branch can be reduced; when the temperature response of the load branch is abnormal, the confidence level of the load forecast can be adjusted. In extreme weather scenarios, this mechanism can adaptively increase the weight of the meteorological branch, enabling the photovoltaic branch's estimation of sudden irradiance changes to be corrected in a timely manner, thereby improving the robustness of the model.

[0062] The probabilistic prediction results are output in Gaussian distribution parameterized form, including the predicted mean and standard deviation of the photovoltaic output and bus load predictions. The predicted mean represents the most likely estimate of the photovoltaic output or load at a future time, serving as a point prediction result; the prediction standard deviation represents the uncertainty of the model's prediction, with a larger standard deviation indicating greater uncertainty. Through Gaussian distribution parameterized output, the model can provide complete probability distribution information for subsequent risk quantification, including predicted values ​​and confidence intervals for each quantile, used to calculate the probability of exceeding the upper threshold load limit and the lower threshold photovoltaic load limit.

[0063] S30: Analyze the probabilistic prediction results, construct a comprehensive risk index accordingly, and output a kilometer-level gridded risk heat map as the risk prediction result based on the comprehensive risk index.

[0064] Finally, the obtained probabilistic prediction results are analyzed, and a comprehensive risk index is constructed accordingly. The comprehensive risk index is a quantitative indicator that comprehensively reflects the degree of supply guarantee risk of the prediction unit within a specific future time window. It represents the probability and severity of the power supply gap faced by the prediction unit due to insufficient photovoltaic output or overload, and is used to rank and classify the risks of each prediction unit.

[0065] The comprehensive risk index is constructed by considering three dimensions: the probability of exceeding the upper load threshold, the probability of exceeding the lower photovoltaic threshold, and the magnitude of the exceeding of the predicted mean. The probability of exceeding the upper load threshold refers to the probability that the predicted load value calculated based on the load forecast distribution exceeds the preset upper load threshold, reflecting the supply guarantee risk caused by excessive electricity demand. The probability of exceeding the lower photovoltaic threshold refers to the probability that the predicted photovoltaic output value calculated based on the photovoltaic output forecast distribution is lower than the preset lower photovoltaic threshold, reflecting the supply guarantee risk caused by insufficient photovoltaic power generation capacity. The magnitude of the exceeding of the predicted mean is quantified by the difference between the predicted load mean and the upper load threshold, and the difference between the lower photovoltaic threshold and the predicted photovoltaic output mean, reflecting the degree of expected shortfall when the risk occurs. The comprehensive risk index is obtained by weighting and integrating these three dimensions.

[0066] Specifically, the probabilistic prediction results are analyzed, and a comprehensive risk index is constructed accordingly, including: Based on the probabilistic prediction results output in Gaussian distribution parameterized form, an approximate cumulative distribution function is obtained analytically; Based on the approximate cumulative distribution function, the probability of exceeding the upper threshold of load and the probability of exceeding the lower threshold of photovoltaic power are calculated respectively, and then weighted and fused to obtain the risk probability. The probability of exceeding the upper threshold of load is the probability that the predicted load value exceeds the preset upper limit threshold, and the probability of exceeding the lower threshold of photovoltaic power is the probability that the predicted photovoltaic output value is lower than the preset lower limit threshold of photovoltaic power. Based on the predicted mean of the probabilistic prediction results, the photovoltaic deficit over-limit and load over-limit are calculated, wherein the photovoltaic deficit over-limit is the positive part of the difference between the photovoltaic lower limit threshold and the photovoltaic output predicted mean, and the load over-limit is the positive part of the difference between the load predicted mean and the load upper limit threshold. A comprehensive risk index is constructed by combining the risk probability, the photovoltaic deficit exceeding the limit, and the load exceeding the limit.

[0067] First, based on the probabilistic prediction results output in Gaussian distribution parameterized form, an approximate cumulative distribution function is analytically obtained. The photovoltaic power output and load prediction results output by the spatiotemporal risk-constrained prediction model are both presented in Gaussian distribution form, meaning that the predicted value at each prediction time step follows a normal distribution with a mean of μ and a standard deviation of σ. Based on this Gaussian distribution, the probability of exceeding any threshold can be analytically calculated using the cumulative distribution function of the standard normal distribution. Specifically, for load prediction, the cumulative distribution function represents the probability that the predicted load value is less than or equal to a certain value; for photovoltaic power output prediction, the cumulative distribution function represents the probability that the predicted photovoltaic power output value is less than or equal to a certain value. Through this cumulative distribution function, the uncertainty of the prediction can be quantified into specific probability values.

[0068] Secondly, based on the approximate cumulative distribution function, the probability of exceeding the upper load threshold and the probability of exceeding the lower photovoltaic threshold are calculated separately, and then weighted and fused to obtain the overall risk probability. The probability of exceeding the upper load threshold is the probability that the predicted load value exceeds the preset upper load threshold, equal to 1 minus the value of the cumulative distribution function at the upper load threshold. When the mean of the predicted load is higher than the upper threshold, this probability is close to 0.5 or higher; when the mean of the predicted load is lower than the upper threshold and the standard deviation is small, this probability is close to 0. The probability of exceeding the lower photovoltaic threshold is the probability that the predicted photovoltaic output value is lower than the preset lower photovoltaic threshold, equal to the value of the cumulative distribution function at the lower photovoltaic threshold. When the mean of the predicted photovoltaic output is lower than the lower threshold, this probability is close to 0.5 or higher; when the mean of the predicted photovoltaic output is higher than the lower threshold and the standard deviation is small, this probability is close to 0. The two exceedance probabilities are weighted and fused according to preset weights to obtain the comprehensive risk probability. The weighting coefficients can be set according to the importance the business places on load risk and photovoltaic risk; for example, the weight for load risk can be set to 0.5, and the weight for photovoltaic risk can be set to 0.5.

[0069] Furthermore, based on the predicted mean of the probabilistic forecast results, the over-limit magnitude of photovoltaic (PV) deficit and the over-limit magnitude of load are calculated. The PV deficit over-limit magnitude is defined as the positive part of the difference between the lower limit threshold and the predicted average PV output. That is, when the predicted average PV output is lower than the lower limit threshold, the over-limit magnitude equals the lower limit threshold minus the predicted average; when the predicted average PV output is higher than or equal to the lower limit threshold, the over-limit magnitude is 0. The load over-limit magnitude is defined as the positive part of the difference between the predicted average load and the upper limit threshold. That is, when the predicted average load is higher than the upper limit threshold, the over-limit magnitude equals the predicted average minus the upper limit threshold; when the predicted average load is lower than or equal to the upper limit threshold, the over-limit magnitude is 0. The over-limit magnitude reflects the severity of the expected deficit when the risk occurs; the larger the magnitude, the more severe the impact once an over-limit event occurs.

[0070] Furthermore, a comprehensive risk index is constructed by combining the risk probability, the extent of photovoltaic deficit / over-limit, and the extent of load over-limit. The comprehensive risk index can be calculated using a weighted summation method, multiplying the risk probability, photovoltaic deficit / over-limit, and load over-limit components by their respective weighting coefficients and then summing them. The weighting coefficients can be set according to the business's preference for the probability and severity of risk occurrence; for example, the risk probability weight could be set to 0.4, the photovoltaic deficit / over-limit weight to 0.3, and the load over-limit weight to 0.3. A higher comprehensive risk index indicates a greater supply guarantee risk for the forecasting unit, requiring priority attention and intervention.

[0071] By integrating these multiple dimensions, the comprehensive risk index considers both the likelihood of a risk occurring and the expected impact of such a risk, making the risk assessment results more comprehensive and accurate.

[0072] Furthermore, a kilometer-scale gridded risk heat map is output based on the comprehensive risk index. This kilometer-scale gridded risk heat map is a two-dimensional color image formed by mapping the comprehensive risk index of each prediction unit onto a spatial grid according to its geographical coordinates. Different colors represent different risk levels; for example, red represents high risk, orange represents medium-high risk, yellow represents medium risk, and blue represents low risk. Through this heat map, dispatchers can intuitively identify the distribution and range of high-risk areas, quickly locate power grid areas requiring key attention and early intervention, and thus formulate targeted supply protection measures, such as adjusting generation plans, optimizing load allocation, or activating emergency power sources. The risk prediction results are used for power dispatching decisions.

[0073] Specifically, a kilometer-scale gridded risk heat map is output as the risk prediction result based on the comprehensive risk index, including: Based on preset grading thresholds, the comprehensive risk index is mapped to grading early warning signals; By traversing multiple prediction units, the corresponding comprehensive risk index is mapped to a kilometer-level grid using a spatial interpolation method based on electrical distance, forming a two-dimensional spatial risk matrix; Based on the graded early warning signals, the two-dimensional spatial risk matrix is ​​rendered using color coding, and a kilometer-level gridded risk heat map is output.

[0074] First, based on preset risk thresholds, the comprehensive risk index is mapped to tiered early warning signals. These preset thresholds are multiple critical values ​​set according to the risk level classification standards in power supply security operations. For example, a blue warning threshold of 0.2, a yellow warning threshold of 0.5, and an orange warning threshold of 0.8 can be set. When the comprehensive risk index is less than or equal to 0.2, it is mapped to a blue warning signal, indicating low risk; when the comprehensive risk index is greater than 0.2 and less than or equal to 0.5, it is mapped to a yellow warning signal, indicating some risk requiring attention; when the comprehensive risk index is greater than 0.5 and less than or equal to 0.8, it is mapped to an orange warning signal, indicating high risk requiring advance preparation; and when the comprehensive risk index is greater than 0.8, it is mapped to a red warning signal, indicating extremely high risk requiring immediate intervention. Through these tiered early warning signals, dispatchers can quickly understand the risk level of each area, facilitating tiered responses.

[0075] Secondly, multiple prediction units are traversed, and the corresponding comprehensive risk index is mapped to a kilometer-level grid using a spatial interpolation method based on electrical distance, forming a two-dimensional spatial risk matrix. Prediction units are discrete spatial locations; for example, each substation or photovoltaic power station corresponds to one prediction unit, and its comprehensive risk index only represents the risk value at that location. To obtain a continuous spatial risk distribution, the risk index of discrete locations needs to be extended to the entire kilometer-level grid using a spatial interpolation method. The spatial interpolation method based on electrical distance uses the electrical distance in the power grid topology as the weighting basis. The electrical distance is determined by line impedance, admittance matrix, or voltage sensitivity between nodes. Specifically, the electrical distance is calculated as follows: based on the power grid node admittance matrix, the equivalent impedance magnitude between nodes is calculated; or based on the voltage sensitivity coefficient matrix, the degree of voltage mutual influence between nodes is calculated. Specifically, for any two prediction units corresponding to power grid nodes i and j, the electrical distance is equal to the equivalent impedance magnitude between node i and node j. A larger impedance magnitude indicates a greater electrical distance and a smaller voltage mutual influence between the two nodes. When a complete admittance matrix cannot be obtained, the geographical distance multiplied by the line tortuosity coefficient can be used as an approximate estimate of the electrical distance.

[0076] Specifically, for any kilometer-level grid point, several surrounding prediction cells are selected, and the electrical distance from the grid point to each prediction cell is calculated. Using the reciprocal of the electrical distance as the weight, the comprehensive risk index of each prediction cell is weighted and averaged to obtain the risk value of the grid point. This process is repeated for all kilometer-level grid points to form a two-dimensional spatial risk matrix.

[0077] Furthermore, based on the tiered early warning signals, a two-dimensional spatial risk matrix is ​​rendered using color coding to output a kilometer-level gridded risk heat map. Color coding is the process of mapping risk levels to different colors; for example, blue corresponds to low risk, yellow to medium risk, orange to medium-high risk, and red to high risk. For each grid point in the two-dimensional spatial risk matrix, a corresponding color value is assigned according to its risk level range, forming a color image. This image uses a map as a base map, overlaid with color layers, to intuitively display the spatial distribution characteristics of the risk. Through the kilometer-level gridded risk heat map, dispatchers can easily identify the geographical location and extent of high-risk areas. For example, if multiple consecutive grids in a certain area are displayed in red, it indicates that the supply guarantee risk in that area is high and requires focused defense. This kilometer-level gridded risk heat map can be output to the dispatch system as a risk prediction result, supporting early warning issuance and dispatch auxiliary decision-making.

[0078] In summary, the embodiments of this application have at least the following technical effects: This invention first acquires multi-source raw data from prediction units and constructs a standard fusion feature vector, integrating heterogeneous data from meteorological observations, remote sensing data, numerical weather prediction, photovoltaic SCADA, bus load, and geographic topology into a unified feature space. This overcomes the limitations of single data sources in characterizing spatial differences and abrupt changes. Second, combined with a pre-trained spatiotemporal risk-constrained prediction model, it outputs probabilistic prediction results encompassing future multi-step photovoltaic output and bus load, achieving a quantitative expression of prediction uncertainty and improving prediction reliability and robustness under extreme weather scenarios. Third, by analyzing the probabilistic prediction results, a comprehensive risk index is constructed, comprehensively considering the probability of exceeding the upper threshold load limit, the probability of exceeding the lower threshold photovoltaic load limit, and the magnitude of the predicted mean exceeding the limit, forming a multi-dimensional fusion risk quantification indicator. Finally, based on the comprehensive risk index, a kilometer-level gridded risk heat map is output, achieving gridded, refined early warning of supply security risks, supporting dispatchers in identifying high-risk areas in advance and formulating targeted supply security measures.

[0079] In summary, this invention solves the problems of insufficient multi-source heterogeneous data fusion capability, low prediction accuracy in extreme weather scenarios, and lack of uncertainty quantification and gridded risk assessment mechanisms in existing technologies.

[0080] Example 2, as Figure 2 As shown, based on the same inventive concept as the power supply risk prediction method based on multi-source data fusion provided in Embodiment 1, this embodiment of the invention also provides a power supply risk prediction system based on multi-source data fusion, including: The feature construction module 11 is used to acquire multi-source raw data of the prediction unit and construct a standard fusion feature vector based on the multi-source raw data; The probabilistic prediction module 12 is used to combine the standard fusion feature vector with the pre-trained spatiotemporal risk constraint prediction model to output a probabilistic prediction result that includes the photovoltaic power output prediction value and the bus load prediction value for multiple future steps. The risk analysis output module 13 is used to analyze the probabilistic prediction results, construct a comprehensive risk index, and output a kilometer-level gridded risk heat map as the risk prediction result based on the comprehensive risk index.

[0081] Specifically, the feature construction module 11 is used for: Acquire multi-source raw data for the prediction unit, and construct a standard fusion feature vector based on the multi-source raw data, including: The multi-source raw data is obtained through multiple data sources, wherein the multi-source raw data includes at least meteorological observation data, remote sensing data, numerical weather prediction data, photovoltaic power generation operation data, bus load data, and power grid geographic topology data. The multi-source raw data is subjected to spatiotemporal alignment, missing value imputation, outlier detection and normalization, and then concatenated into the standard fusion feature vector. The standard fusion feature vector includes meteorological element components, photovoltaic power output historical sequence components, load historical sequence components, static geographic equipment feature components, and calendar feature components.

[0082] Specifically, the probabilistic prediction module 12 is used for: Combining the aforementioned standard fusion feature vector with the pre-trained spatiotemporal risk-constrained prediction model, the output is a probabilistic prediction result containing multi-step future photovoltaic power output and bus load predictions, including: By using a priori temporal encoder and spatial feature aggregator, the standard fused feature vector is spatiotemporally encoded to obtain a comprehensive spatiotemporal representation vector; The spatiotemporal risk-constrained prediction model is used to perform probabilistic prediction of the comprehensive spatiotemporal representation vector based on a three-branch decoding channel, wherein the three-branch decoding channel includes a meteorological prediction branch, a photovoltaic power output prediction branch, and a load prediction branch. The outputs of the three-branch decoding channels are subjected to inter-modal adaptive fusion to obtain probabilistic prediction results of photovoltaic power output prediction and bus load prediction. The probabilistic prediction results are output in Gaussian distribution parameterized form, including the predicted mean and predicted standard deviation of the photovoltaic power output prediction value and the bus load prediction value.

[0083] Specifically, the standard fused feature vector is spatiotemporally encoded using a priori temporal encoder and spatial feature aggregator to obtain a comprehensive spatiotemporal representation vector, including: A temporal encoder is used to extract temporal features from the standard fused feature vector to obtain temporal feature embeddings. Using the adjacency matrix constructed based on power grid geographic topology data as a constraint, the spatial feature embedding is obtained by aggregating spatial dimension features of the temporal feature embedding through the spatial feature aggregator; The temporal feature embedding, the spatial feature embedding, and the standard fused feature vector are concatenated to obtain a comprehensive spatiotemporal representation; The temporal encoder is implemented based on any one of bidirectional LSTM, temporal convolutional network, and temporal Transformer, and the spatial feature aggregator is implemented based on any one of graph convolution and graph attention.

[0084] Specifically, the spatiotemporal risk constraint prediction model is used to perform probabilistic prediction of the integrated spatiotemporal representation vector based on a three-branch decoding channel, including: The meteorological forecasting branch receives the integrated spatiotemporal representation vector, decodes and reconstructs it, and outputs the probabilistic forecast results of meteorological elements in the form of Gaussian distribution parameterization for the next multiple steps. The load forecasting branch receives the integrated spatiotemporal representation vector, decodes and reconstructs it, and outputs the probabilistic prediction result of the bus load in the parameterized form of Gaussian distribution for the next multiple steps. The photovoltaic output prediction branch receives the integrated spatiotemporal representation vector and the probabilistic prediction results of the meteorological elements, decodes and reconstructs them, and outputs the probabilistic prediction results of photovoltaic output in the form of Gaussian distribution parameterization for the next many steps.

[0085] The training and construction process of the spatiotemporal risk constraint prediction model includes: Acquire training sample data, wherein the training sample data includes a sample standard fusion feature vector and a corresponding supervision target vector, and the supervision target vector includes the true values ​​of meteorological elements, photovoltaic power output, and bus load; The three-branch decoding channel is constructed based on a deterministic neural network, wherein the decoding output layer of each branch is set as a distributed parameterized output head, the distributed parameterized output head includes a mean output subheader and a standard deviation output subheader, and the standard deviation output subheader is constrained to have a positive output value by an activation function; The time encoder, the spatial feature aggregator, the three-branch decoding channel, and the cross-modal attention mechanism are integrated to form the overall network structure of the spatiotemporal risk constraint prediction model. The spatiotemporal risk constraint prediction model is trained in a supervised manner using the training sample data and a hybrid loss function. In this cross-modal attention mechanism, the query vector comes from the feature representation of the meteorological forecasting branch, and the key vector and value vector come from the feature representations of the photovoltaic power output forecasting branch and the load forecasting branch. The adaptive allocation of intermodal weights is achieved through scaling dot product attention calculation.

[0086] Specifically, the hybrid loss function is a weighted combination of three terms: mean squared error, quantile loss, and risk penalty, including: The mean square error between the predicted mean and the true value is defined as the mean square error term. Based on the preset list of quantiles, the predicted value of each quantile is calculated by combining the predicted mean and the predicted standard deviation, and the average of the loss between the predicted quantile value and the true quantile value is defined as the quantile loss term. The risk penalty term is defined by combining a preset relative error threshold and an activation function. The risk penalty term is used to generate a penalty gradient through the activation function when the prediction deviation exceeds the relative error threshold.

[0087] Specifically, the risk analysis output module 13 is used for: Analyzing the probabilistic prediction results, a comprehensive risk index is constructed, including: Based on the probabilistic prediction results output in Gaussian distribution parameterized form, an approximate cumulative distribution function is obtained analytically; Based on the approximate cumulative distribution function, the probability of exceeding the upper threshold of load and the probability of exceeding the lower threshold of photovoltaic power are calculated respectively, and then weighted and fused to obtain the risk probability. The probability of exceeding the upper threshold of load is the probability that the predicted load value exceeds the preset upper limit threshold, and the probability of exceeding the lower threshold of photovoltaic power is the probability that the predicted photovoltaic output value is lower than the preset lower limit threshold of photovoltaic power. Based on the predicted mean of the probabilistic prediction results, the photovoltaic deficit over-limit and load over-limit are calculated, wherein the photovoltaic deficit over-limit is the positive part of the difference between the photovoltaic lower limit threshold and the photovoltaic output predicted mean, and the load over-limit is the positive part of the difference between the load predicted mean and the load upper limit threshold. A comprehensive risk index is constructed by combining the risk probability, the photovoltaic deficit exceeding the limit, and the load exceeding the limit.

[0088] Specifically, a kilometer-scale gridded risk heat map is output as the risk prediction result based on the comprehensive risk index, including: Based on preset grading thresholds, the comprehensive risk index is mapped to grading early warning signals; By traversing multiple prediction units, the corresponding comprehensive risk index is mapped to a kilometer-level grid using a spatial interpolation method based on electrical distance, forming a two-dimensional spatial risk matrix; Based on the graded early warning signals, the two-dimensional spatial risk matrix is ​​rendered using color coding, and a kilometer-level gridded risk heat map is output.

Claims

1. A power supply risk prediction method based on multi-source data fusion, characterized in that, include: Obtain multi-source raw data for the prediction unit, and construct a standard fusion feature vector based on the multi-source raw data; Combining the standard fusion feature vector with the pre-trained spatiotemporal risk constraint prediction model, the output is a probabilistic prediction result containing the photovoltaic power output prediction value and the bus load prediction value for multiple future steps. Analyze the probabilistic prediction results, construct a comprehensive risk index, and output a kilometer-level gridded risk heat map as the risk prediction result based on the comprehensive risk index.

2. The power supply risk prediction method based on multi-source data fusion as described in claim 1, characterized in that, Acquire multi-source raw data for the prediction unit, and construct a standard fusion feature vector based on the multi-source raw data, including: The multi-source raw data is obtained through multiple data sources, wherein the multi-source raw data includes at least meteorological observation data, remote sensing data, numerical weather prediction data, photovoltaic power generation operation data, bus load data, and power grid geographic topology data. The multi-source raw data is subjected to spatiotemporal alignment, missing value imputation, outlier detection and normalization, and then concatenated into the standard fusion feature vector. The standard fusion feature vector includes meteorological element components, photovoltaic power output historical sequence components, load historical sequence components, static geographic equipment feature components, and calendar feature components.

3. The power supply risk prediction method based on multi-source data fusion as described in claim 1, characterized in that, Combining the aforementioned standard fusion feature vector with the pre-trained spatiotemporal risk-constrained prediction model, the output is a probabilistic prediction result containing multi-step future photovoltaic power output and bus load predictions, including: By using a priori temporal encoder and spatial feature aggregator, the standard fused feature vector is spatiotemporally encoded to obtain a comprehensive spatiotemporal representation vector; The spatiotemporal risk-constrained prediction model is used to perform probabilistic prediction of the comprehensive spatiotemporal representation vector based on a three-branch decoding channel, wherein the three-branch decoding channel includes a meteorological prediction branch, a photovoltaic power output prediction branch, and a load prediction branch. The outputs of the three-branch decoding channels are subjected to inter-modal adaptive fusion to obtain probabilistic prediction results of photovoltaic power output prediction and bus load prediction. The probabilistic prediction results are output in Gaussian distribution parameterized form, including the predicted mean and predicted standard deviation of the photovoltaic power output prediction value and the bus load prediction value.

4. The power supply risk prediction method based on multi-source data fusion as described in claim 3, characterized in that, By using a priori temporal encoder and spatial feature aggregator, the standard fused feature vector is spatiotemporally encoded to obtain a comprehensive spatiotemporal representation vector, including: A temporal encoder is used to extract temporal features from the standard fused feature vector to obtain temporal feature embeddings. Using the adjacency matrix constructed based on power grid geographic topology data as a constraint, the spatial feature embedding is obtained by aggregating spatial dimension features of the temporal feature embedding through the spatial feature aggregator; The temporal feature embedding, the spatial feature embedding, and the standard fused feature vector are concatenated to obtain a comprehensive spatiotemporal representation; The temporal encoder is implemented based on any one of bidirectional LSTM, temporal convolutional network, and temporal Transformer, and the spatial feature aggregator is implemented based on any one of graph convolution and graph attention.

5. The power supply risk prediction method based on multi-source data fusion as described in claim 3, characterized in that, Using the aforementioned spatiotemporal risk-constrained prediction model, probabilistic prediction of the comprehensive spatiotemporal representation vector based on a three-branch decoding channel is performed, including: The meteorological forecasting branch receives the integrated spatiotemporal representation vector, decodes and reconstructs it, and outputs the probabilistic forecast results of meteorological elements in the form of Gaussian distribution parameterization for the next multiple steps. The load forecasting branch receives the integrated spatiotemporal representation vector, decodes and reconstructs it, and outputs the probabilistic prediction result of the bus load in the parameterized form of Gaussian distribution for the next multiple steps. The photovoltaic output prediction branch receives the integrated spatiotemporal representation vector and the probabilistic prediction results of the meteorological elements, decodes and reconstructs them, and outputs the probabilistic prediction results of photovoltaic output in the form of Gaussian distribution parameterization for the next many steps.

6. The power supply risk prediction method based on multi-source data fusion as described in claim 3, characterized in that, The training and construction process of the spatiotemporal risk constraint prediction model includes: Acquire training sample data, wherein the training sample data includes a sample standard fusion feature vector and a corresponding supervision target vector, and the supervision target vector includes the true values ​​of meteorological elements, photovoltaic power output, and bus load; The three-branch decoding channel is constructed based on a deterministic neural network, wherein the decoding output layer of each branch is set as a distributed parameterized output head, the distributed parameterized output head includes a mean output subheader and a standard deviation output subheader, and the standard deviation output subheader is constrained to have a positive output value by an activation function; The time encoder, the spatial feature aggregator, the three-branch decoding channel, and the cross-modal attention mechanism are integrated to form the overall network structure of the spatiotemporal risk constraint prediction model. The spatiotemporal risk constraint prediction model is trained in a supervised manner using the training sample data and a hybrid loss function. In this cross-modal attention mechanism, the query vector comes from the feature representation of the meteorological forecasting branch, and the key vector and value vector come from the feature representations of the photovoltaic power output forecasting branch and the load forecasting branch. The adaptive allocation of intermodal weights is achieved through scaling dot product attention calculation.

7. The power supply risk prediction method based on multi-source data fusion as described in claim 6, characterized in that, The hybrid loss function is a weighted combination of three terms: mean squared error, quantile loss, and risk penalty. The mean square error between the predicted mean and the true value is defined as the mean square error term. Based on the preset list of quantiles, the predicted value of each quantile is calculated by combining the predicted mean and the predicted standard deviation, and the average of the loss between the predicted quantile value and the true quantile value is defined as the quantile loss term. The risk penalty term is defined by combining a preset relative error threshold and an activation function. The risk penalty term is used to generate a penalty gradient through the activation function when the prediction deviation exceeds the relative error threshold.

8. The power supply risk prediction method based on multi-source data fusion as described in claim 1, characterized in that, Analyzing the probabilistic prediction results, a comprehensive risk index is constructed, including: Based on the probabilistic prediction results output in Gaussian distribution parameterized form, an approximate cumulative distribution function is obtained analytically; Based on the approximate cumulative distribution function, the probability of exceeding the upper threshold of load and the probability of exceeding the lower threshold of photovoltaic power are calculated respectively, and then weighted and fused to obtain the risk probability. The probability of exceeding the upper threshold of load is the probability that the predicted load value exceeds the preset upper limit threshold, and the probability of exceeding the lower threshold of photovoltaic power is the probability that the predicted photovoltaic output value is lower than the preset lower limit threshold of photovoltaic power. Based on the predicted mean of the probabilistic prediction results, the photovoltaic deficit over-limit and load over-limit are calculated, wherein the photovoltaic deficit over-limit is the positive part of the difference between the photovoltaic lower limit threshold and the photovoltaic output predicted mean, and the load over-limit is the positive part of the difference between the load predicted mean and the load upper limit threshold. A comprehensive risk index is constructed by combining the risk probability, the photovoltaic deficit exceeding the limit, and the load exceeding the limit.

9. The power supply risk prediction method based on multi-source data fusion as described in claim 1, characterized in that, Based on the comprehensive risk index, a kilometer-scale gridded risk heat map is output as the risk prediction result, including: Based on preset grading thresholds, the comprehensive risk index is mapped to grading early warning signals; By traversing multiple prediction units, the corresponding comprehensive risk index is mapped to a kilometer-level grid using a spatial interpolation method based on electrical distance, forming a two-dimensional spatial risk matrix; Based on the graded early warning signals, the two-dimensional spatial risk matrix is ​​rendered using color coding, and a kilometer-level gridded risk heat map is output.

10. A power supply risk prediction system based on multi-source data fusion, characterized in that, The method for performing the power supply risk prediction method based on multi-source data fusion as described in any one of claims 1-9 includes: The feature construction module is used to acquire multi-source raw data of the prediction unit and construct a standard fusion feature vector based on the multi-source raw data; The probabilistic prediction module is used to combine the standard fused feature vector with the pre-trained spatiotemporal risk-constrained prediction model to output probabilistic prediction results that include the photovoltaic power output prediction value and the bus load prediction value for multiple future steps. The risk analysis output module is used to analyze the probabilistic prediction results, construct a comprehensive risk index, and output a kilometer-level gridded risk heat map as the risk prediction result based on the comprehensive risk index.

Images