Time sequence probability prediction method for fusion of two-stage space-time diagram network and multi-source information

Abstract

The invention discloses a two-stage space-time diagram network and multi-source information fusion time sequence probability prediction method and device, and relates to the technical field of artificial intelligence and big data analysis. The method comprises the following steps: acquiring historical wind speed sequence data and corresponding target variable sequence data; preprocessing the data, and constructing a training sample according to a preset time sliding window; constructing a time sequence probability prediction framework of the two-stage space-time diagram network and multi-source information fusion; based on the fluctuation correlation of the historical sequence data, constructing an adjacent matrix between nodes; inputting the training sample and the adjacent matrix between the nodes into a wind speed prediction model, training the model through a designed threshold perception loss function of a wind speed-power nonlinear relationship, and outputting a multi-node wind speed prediction probability of a first stage; and inputting the multi-node wind speed prediction probability and the multi-source environmental factor data into the gradient boosting tree model for second-stage prediction, and outputting a multi-node wind power prediction result. According to the invention, prediction errors can be reduced.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and big data analysis technology, and in particular to a temporal probability prediction method and apparatus that fuses a two-stage spatiotemporal graph network with multi-source information. Background Technology

[0002] Time-series data forecasting is widely used in wind power forecasting, weather forecasting, traffic flow analysis, energy load management, and financial risk control. This technology uses historical observation sequences to infer the changing trends of target variables over future periods. With the high proportion of renewable energy connected to the grid and the widespread adoption of multi-source heterogeneous sensing devices, forecasting demand is shifting from "single deterministic point values" to "probabilistic forecasts that include quantified uncertainty," in order to support robustness in scheduling optimization, energy storage configuration, and risk management.

[0003] The core idea of ​​time series data prediction is usually to extrapolate short-term target variables based on historical measurement sequences using time series modeling methods. However, in practical applications, the spatiotemporal coupling of multiple nodes and sources often significantly increases the difficulty of prediction, prompting researchers to continuously explore composite models that combine spatial and temporal features. Among existing similar technologies, a typical approach is based on single-stage end-to-end deep time series models such as recurrent neural networks, long short-term memory networks, gated recurrent units, or hybrid networks combining convolutional and recurrent structures. These models can capture temporal dependencies and local changes to some extent, but they lack sufficient spatial modeling capabilities when facing multi-node systems with significant spatial correlations. In addition, some works introduce graph neural networks or graph convolutional networks to model the static spatial relationships between nodes and combine them with Transformer structures using self-attention mechanisms to characterize temporal dependencies. However, conventional self-attention has high computational and storage overhead in multi-node long sequence prediction, resulting in high engineering deployment costs and making it unsuitable for scenarios with high real-time requirements.

[0004] To address the need for joint spatial and temporal modeling, existing research has proposed a "two-stage prediction framework." The first stage predicts easily modelable intermediate variables (such as wind speed and traffic flow), while the second stage uses these predictions, combined with external environmental factors (such as temperature, humidity, and air pressure), to infer the final target variable through regression or ensemble learning models. This approach effectively reduces the complexity of direct end-to-end modeling and allows for the selection of more targeted models at different stages. However, such frameworks still have several shortcomings in their specific structural implementation: First, the predictions in the first stage are mostly point values, lacking uncertainty characterization and probability information output, which makes it impossible to effectively transmit and quantify the prediction error in the second stage; Second, the multi-source fusion of environmental factors is mostly achieved by simple vector concatenation or feature stacking, lacking structured attention mechanisms or importance weighting, and failing to fully highlight the key features that make significant contributions to the prediction, thus easily introducing interference when there is noise or outliers in the data; Third, the probability prediction module in the second stage relies on single quantile regression or ex-post interval construction methods based on residuals in many existing technologies. This approach is prone to unstable interval coverage or excessively large or small interval widths in practical applications, which reduces the reliability and practicality of the prediction.

[0005] Another existing approach combines clustering with local modeling. This involves clustering multi-node data and then using long short-term memory networks, convolutional neural networks, gated recurrent units, or other short-term prediction models to model the data step-by-step within each cluster. While this approach alleviates the challenges of spatiotemporal feature extraction to some extent, it suffers from insufficient adaptability. When wind field configurations, environmental conditions, or inter-node correlations change, re-clustering and retraining are required to adapt to the new structural relationships, increasing update and maintenance costs and reducing the stability and continuity of online predictions. Some static graph-structured spatiotemporal models also experience performance degradation when facing data drift because the adjacency matrix and spatial topology cannot be adaptively updated, leading to distortions in spatial feature extraction and temporal dynamic characterization.

[0006] In summary, the common problems of existing technologies mainly revolve around three aspects: insufficient joint characterization of spatial correlation and temporal dependence, limited ability to predict probabilistically and transmit uncertainty, and simple multi-source information fusion mechanisms with weak anti-interference capabilities. These shortcomings lead to a decrease in the accuracy of prediction models in highly dynamic, multi-node, and multi-source heterogeneous real-world application scenarios, making it difficult to provide highly reliable decision support for scheduling optimization and risk control. Summary of the Invention

[0007] To address the technical problems of insufficient prediction accuracy, weak uncertainty characterization, and low efficiency of multi-source information fusion in existing technologies under highly dynamic, multi-node, and multi-source heterogeneous scenarios, this invention provides a temporal probability prediction method and apparatus that integrates a two-stage spatiotemporal graph network with multi-source information fusion. The technical solution is as follows:

[0008] On the one hand, a temporal probability prediction method based on the fusion of a two-stage spatiotemporal graph network and multi-source information is provided. This method is implemented by a temporal probability prediction device based on the fusion of a two-stage spatiotemporal graph network and multi-source information, and includes: S1. Obtain historical wind speed sequence data and corresponding real wind power sequence data from the multi-node sensor network of the short-term wind power prediction system of the wind farm; preprocess the historical wind speed sequence data and corresponding real wind power sequence data, and construct training samples according to a preset time sliding window; S2. A temporal probabilistic prediction framework for fusion of a two-stage spatiotemporal graph network and multi-source information is constructed; the framework includes: a wind speed prediction model and a gradient boosting tree model; S3. Based on the fluctuation correlation of historical sequence data, construct the adjacency matrix between nodes; input the training samples and the adjacency matrix between nodes into the wind speed prediction model, train the wind speed prediction model through the threshold perception loss function of the designed wind speed-power nonlinear relationship, obtain the trained wind speed prediction model, predict the wind speed probability of real-time input based on the trained wind speed prediction model, and output the multi-node wind speed prediction probability of the first stage. S4. Input the multi-node wind speed prediction probability and the acquired multi-source environmental factor data into the gradient boosting tree model for the second stage of prediction, and output the second stage multi-node wind power prediction results.

[0009] On the other hand, a temporal probability prediction device based on the fusion of a two-stage spatiotemporal graph network and multi-source information is provided. This device is applied to the temporal probability prediction method based on the fusion of a two-stage spatiotemporal graph network and multi-source information. The device includes: The acquisition unit is used to acquire historical wind speed sequence data and corresponding real wind power sequence data from the multi-node sensor network of the short-term wind power prediction system of the wind farm; preprocess the historical wind speed sequence data and corresponding real wind power sequence data, and construct training samples according to a preset time sliding window. A building block is used to construct a temporal probabilistic prediction framework that integrates a two-stage spatiotemporal graph network with multi-source information fusion; the framework includes a wind speed prediction model and a gradient boosting tree model. The first prediction unit is used to construct an adjacency matrix between nodes based on the fluctuation correlation of historical sequence data; input the training samples and the adjacency matrix between nodes into the wind speed prediction model, train the wind speed prediction model through the threshold-aware loss function of the designed wind speed-power nonlinear relationship, obtain the trained wind speed prediction model, predict the wind speed probability of real-time input based on the trained wind speed prediction model, and output the multi-node wind speed prediction probability of the first stage. The second prediction unit is used to input the multi-node wind speed prediction probability and the acquired multi-source environmental factor data into the gradient boosting tree model for the second stage prediction, and output the second stage multi-node wind power prediction result.

[0010] On the other hand, a temporal probability prediction device for fusion of a two-stage spatiotemporal graph network and multi-source information is provided. The temporal probability prediction device for fusion of a two-stage spatiotemporal graph network and multi-source information includes: a processor; a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, any one of the methods in the above-described temporal probability prediction method for fusion of a two-stage spatiotemporal graph network and multi-source information is implemented.

[0011] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described temporal probabilistic prediction methods of two-stage spatiotemporal graph networks and multi-source information fusion.

[0012] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention presents a probabilistic prediction method for time-series data by constructing a two-stage spatiotemporal graph network and fusing multi-source information. In the first stage, a dynamically updatable adjacency matrix is ​​constructed based on the correlation of historical data fluctuations. Combined with a graph convolutional network and a spatiotemporal Transformer incorporating an external memory mechanism, this effectively characterizes the complex spatial dependencies in multi-node systems and their dynamic characteristics over time, significantly improving the accuracy and stability of intermediate variable (e.g., wind speed) predictions. The introduction of an external attention mechanism maintains global spatiotemporal dependency modeling capabilities while reducing computational and storage overhead, making this invention more suitable for high-dimensional, multi-node, and long-sequence practical engineering scenarios, and solving the problem of excessively high computational costs associated with traditional self-attention methods. Combined with the temporal feature extraction structure of a multilayer perceptron and the weighted optimization of key business intervals using a threshold-aware loss function, this invention achieves more accurate fitting and generalization in important segments affecting the final goal, effectively reducing the impact of prediction errors on decision-making.

[0013] In the second stage, this invention introduces multi-source environmental factors such as temperature, humidity, and air pressure, and employs a gradient boosting tree model to achieve deep integration of intermediate variables and external factors. This fully leverages the ensemble model's ability to characterize complex nonlinear relationships and feature interactions, improving the accuracy and interpretability of target variable (e.g., wind power) predictions. By optimizing model parameters through grid search and cross-validation, this invention achieves excellent performance across multiple accuracy and probability metrics, ensuring the prediction results possess acceptable accuracy and stability for engineering applications. Furthermore, this invention establishes an uncertainty propagation mechanism between the two stages, converting the probabilistic prediction information from the first stage into usable feature inputs for the downstream stage. This allows the second-stage model to fully perceive the prediction range and error morphology, enhancing its resilience during target mapping and reducing the cumulative amplification of upstream prediction errors.

[0014] The embodiments of this invention can significantly improve the prediction accuracy and stability of multi-node and multi-source heterogeneous time series data; it realizes the effective transmission of probabilistic information from intermediate variables to target variables, enhancing the credibility and resilience of prediction results; it optimizes the balance between coverage and width of interval prediction, providing accurate and reliable decision support for engineering scheduling; it reduces the computational overhead of large-scale spatiotemporal modeling, and has good real-time performance and deployment feasibility; the framework proposed in the embodiments of this invention has strong versatility and can be directly transferred and extended in various spatiotemporal coupled prediction scenarios such as new energy power, traffic management, weather forecasting, and industrial Internet of Things, meeting the needs of different fields for high-precision and high-reliability time series prediction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a temporal probability prediction method that integrates a two-stage spatiotemporal graph network and multi-source information, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a probabilistic prediction method based on a two-stage spatiotemporal graph network and multi-source information fusion provided in an embodiment of the present invention; Figure 3 This is a block diagram of a temporal probability prediction device that fuses a two-stage spatiotemporal graph network with multi-source information, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the structure of a time-series probability prediction device that integrates a two-stage spatiotemporal graph network and multi-source information, provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] This invention provides a temporal probability prediction method based on the fusion of a two-stage spatiotemporal graph network and multi-source information. This method can be implemented using a temporal probability prediction device that integrates a two-stage spatiotemporal graph network and multi-source information, which can be a terminal or a server. Figure 1 The flowchart shown is for a temporal probability prediction method that fuses a two-stage spatiotemporal graph network with multi-source information. The processing flow of this method may include the following steps:

[0023] S1. Obtain historical wind speed sequence data and corresponding real wind power sequence data from the multi-node sensor network of the short-term wind power prediction system of the wind farm; preprocess the historical wind speed sequence data and corresponding real wind power sequence data, and construct training samples according to the preset time sliding window.

[0024] This invention can be applied to multi-source heterogeneous scenarios, which refer to scenarios where the data required for prediction tasks comes from two or more different sources and differs in data type, structure, spatial / temporal resolution, sampling frequency, unit of measurement, timeliness, and quality characteristics (including noise, missing data, and drift). These scenarios require alignment, representation, and fusion within a unified computing framework to support spatiotemporal modeling and temporal probabilistic prediction. Typical multi-source heterogeneous scenarios include new energy and power system scenarios, transportation and travel scenarios, and meteorological and hydrological scenarios.

[0025] In one feasible implementation, the present invention collects historical wind speed sequence data and corresponding target variable sequence data, and simultaneously collects multi-source external information closely related to the prediction target from relevant data sources, such as meteorological data, equipment operating parameters, and operating status data.

[0026] In one feasible implementation, the preprocessing steps of this invention include: time alignment, missing value imputation, outlier cleaning, and normalization. Constructing training samples ensures that the length, step size, and prediction step size of the input sequence meet the timeliness and stability requirements of short-term prediction tasks.

[0027] in, Figure 2 This is a schematic diagram of a probabilistic prediction method based on a two-stage spatiotemporal graph network and multi-source information fusion provided by an embodiment of the present invention. In one feasible implementation, the probabilistic prediction process based on the two-stage spatiotemporal graph network and multi-source information fusion includes two stages; the first stage involves inputting the obtained spatiotemporal wind speed data and the calculated adjacency matrix into the constructed wind speed prediction model, and extracting node features through a spatial feature extraction module; the node features are then input into a temporal feature extraction module, processed by a multilayer perceptron, and outputting multi-node wind speed prediction data; the multi-node wind speed prediction data and the acquired multi-source environmental factor data are then input into a gradient boosting tree model, and the weak learner is trained and its weights are updated to output multi-node wind power prediction results.

[0028] S2. A temporal probabilistic prediction framework for the fusion of a two-stage spatiotemporal graph network and multi-source information is constructed. The framework includes a wind speed prediction model and a gradient boosting tree model.

[0029] The gradient boosting tree model refers to a boosting algorithm where the weak learner only uses decision trees (usually CART regression trees). Its core logic is to train multiple decision trees sequentially, with each tree specifically correcting the prediction error of the preceding model. The negative gradient of the loss function is used as the optimization objective, and finally, the prediction results of all trees are weighted and fused to form a strong learner.

[0030] Optionally, the wind speed prediction model includes: a spatial feature extraction module and a temporal feature extraction module; The spatial feature extraction module consists of a graph convolutional network and a spatial Transformer model with an external attention mechanism. The time feature extraction module consists of a multi-layer sensing mechanism.

[0031] S3. Based on the fluctuation correlation of historical sequence data, construct the adjacency matrix between nodes; input the training samples and the adjacency matrix between nodes into the wind speed prediction model, train the wind speed prediction model through the threshold-aware loss function of the designed wind speed-power nonlinear relationship, obtain the trained wind speed prediction model, predict the real-time input wind speed probability based on the trained wind speed prediction model, and output the multi-node wind speed prediction probability of the first stage.

[0032] In this embodiment of the invention, an adjacency matrix between nodes is constructed based on the fluctuation correlation of each node's historical sequence. A weighted graph structure is generated using metrics such as Pearson or Spearman correlation coefficients, mutual information, or time-delay correlation. Noise edges are suppressed through thresholding or K-nearest neighbor sparsification. As the physical or geographical topology or statistical correlation structure of the system evolves over time, the adjacency matrix is ​​adaptively updated within a moving window to form a dynamic graph representation, thereby improving the adaptability to time-varying spatial relationships. The system refers to a set of target spatiotemporal objects (such as wind farms, wind turbines, and meteorological tower observation networks) containing multiple spatially distributed observation nodes and outputting time series. The system can also refer to multi-node spatiotemporal systems such as power grids, transportation networks, hydrological station networks, industrial IoT, and urban IoT, and is not limited to any specific system type.

[0033] Optionally, the specific implementation process of S3 includes S31-S33: S31. Input the training samples and the adjacency matrix between nodes into the spatial feature extraction module in the wind speed prediction model. Use a graph convolutional network to perform weighted aggregation and feature transformation on the adjacency matrix of the nodes, and output the temporal features of the nodes. In one feasible implementation, the specific process of weighted aggregation and feature transformation of the adjacency matrix of nodes through a graph convolutional network to output the temporal features of nodes is represented by the following formula (1): (1) in, , Degree matrix; Indicates the first Layer node characteristics; Represents the learnable weight matrix; Indicates the activation function; Indicates the first Layer node characteristics.

[0034] S32. Superimpose the node temporal features and location codes and input them into the spatial Transformer model with an external attention mechanism. Calculate the weight matrix of the external attention mechanism through the external attention mechanism. In one feasible implementation, the external attention mechanism calculation process is represented by the following formula (2): (2) in, Represents the query matrix; The key matrix representing external memory; The value matrix representing external memory; Indicates the feature dimension scaling factor; Represents the output function of the external attention mechanism; This represents the normalization function, which transforms the matching score into a weighted probability distribution, ensuring that the sum of all weights is 1.

[0035] S33. Input the weight matrix of the external attention mechanism into the time feature extraction module, perform nonlinear mapping through a multilayer perceptron, and output the multi-node wind speed prediction probability; based on the multi-node wind speed prediction probability, train the wind speed prediction model through the threshold perception loss function of the designed wind speed-power nonlinear relationship to obtain the trained wind speed prediction model; based on the trained wind speed prediction model, predict the multi-node wind speed prediction probability of the first stage. The process of nonlinear mapping through a multilayer perceptron is represented by the following formula (3): (3) in, This represents the output vector at the current time step index t. Represents the MLP mapping function; Indicates from The matrix formed by concatenating the input sequence up to t according to time. Represents a non-linear activation function; This represents the weight matrix of the first layer of the MLP; This represents the weight matrix of the second layer of the MLP; This represents the bias vector of the first layer of the MLP; This represents the bias vector of the second layer of the MLP; This represents the input feature matrix obtained by concatenating the observations of the object under study in chronological order within the time window [t-w+1, t]; t represents the time step index in the prediction sequence.

[0036] Alternatively, the threshold-perceived loss function for the wind speed-power nonlinear relationship is expressed by the following formula (4): (4) in, This represents the total value of the weighted quantile loss function; This represents the weighting coefficient, which is related to the distance between the current sample and the key business threshold. This represents the actual value of the current time step index t; This represents the predicted value at the current time step index t; represents the quantile level; t represents the time step index in the predicted sequence.

[0037] Among them, the threshold-aware loss function of the wind speed-power nonlinear relationship dynamically adjusts the loss weight of different samples according to the relationship between the wind speed prediction value and the preset threshold, thereby improving the prediction accuracy of the model in the key wind speed range.

[0038] The threshold-aware loss function for the nonlinear relationship between wind speed and power is designed based on the physical characteristics of the power curve: when the wind speed exceeds a certain threshold (e.g., 15 m / s), the wind turbine output power tends to saturate, and the impact of wind speed fluctuations on power weakens. Therefore, the tolerance of the wind speed prediction model for high wind speed prediction errors can be appropriately increased. The loss function achieves sample-level weighting by introducing a weight mask mechanism. This loss function is integrated into the backpropagation process during the first stage of wind speed prediction model training. The gradient descent algorithm optimizes the wind speed prediction model parameters, thereby achieving accurate modeling of key wind speed ranges. After training, the wind speed prediction model can be used to efficiently and accurately predict future wind speed sequences and provide high-quality input features for the second stage of wind power probability prediction.

[0039] In one feasible implementation, to effectively propagate upstream prediction uncertainty downstream, this invention designs an uncertainty propagation mechanism based on probabilistic feature splicing between the two stages. The core of this mechanism is to directly use the multidimensional probabilistic features output by the first-stage model, which fully characterize the prediction uncertainty of intermediate variables, as part of the input features of the second-stage gradient boosting tree model. This allows the downstream gradient boosting tree model to explicitly perceive and utilize the reliability information of the upstream prediction.

[0040] In one feasible implementation, the specific process of the uncertainty propagation mechanism based on probabilistic feature splicing includes: the present invention uses a spatiotemporal network model to predict the predicted values ​​at multiple target quantile levels.

[0041] S4. Input the multi-node wind speed prediction probability and the acquired multi-source environmental factor data into the gradient boosting tree model for the second stage of prediction, and output the second stage multi-node wind power prediction results.

[0042] Optionally, multi-source environmental data may include temperature, humidity, and air pressure.

[0043] Optionally, a specific implementation of S4 includes S41-S43: S41. Based on the multi-node wind speed prediction probability input into the pre-trained spatiotemporal network model, the quantiles corresponding to future time steps are predicted based on the pre-selected quantile level set, and the quantile prediction vector is output. The Spatio-Temporal Network Model (PTM) is a modeling method for complex systems that integrates spatial topology with temporal dynamics. Its core idea is to abstract the interactions between entities into a network (nodes represent entities, edges represent relationships) and explicitly model the evolution of these relationships over time. This model is widely used in traffic prediction, pedestrian flow analysis, social networks, and communication networks, and is one of the core technologies for processing spatio-temporal data.

[0044] In one feasible implementation, the pre-selected set of quantile levels is represented as: ,For example, Then, the quantile prediction vector for a future time t, generated by the spatiotemporal network model, can be expressed by the following formula (5): (5) in, Indicates at time step quantile vector; Indicates the corresponding quantile level The predicted value below; Indicates the first quantile level, range of values , It represents the total number of quantiles, i.e., the number of quantiles extracted from the predicted distribution.

[0045] S42. Concatenate the quantile prediction vector with the features extracted from the wind speed prediction model to obtain the concatenated feature vector. In one feasible implementation, during the uncertainty propagation phase, the quantile prediction vector is concatenated with the node features obtained from the first-stage model.

[0046] S43. The spliced ​​feature vector is fused with multi-source environmental factor data to obtain a fused feature vector; the fused feature vector is input into the gradient boosting tree model for the second stage prediction, and the second stage multi-node wind power prediction result is output.

[0047] In one feasible implementation, the concatenated feature vectors are further integrated with multi-source environmental factors, including temperature, humidity, and air pressure, to form the complete input feature set of the second-stage target variable mapping model. In this way, the input of the gradient boosting tree model for the second stage includes structured uncertainty information from upstream sources. During training, the gradient boosting tree model can learn the complex relationship between these quantile features and the final target variable, wind power. Therefore, when making point or quantile predictions, it can fully consider the confidence level of upstream predictions, achieving explicit cross-stage propagation and utilization of uncertainty, ultimately improving the calibration quality and reliability of the overall probability prediction.

[0048] Optionally, the process of inputting the fused feature vectors into the gradient boosting tree model for the second-stage prediction is represented by the following formula (6): (6) in, This represents the predicted value of the target variable at time t; Indicates the number of weak learners; Indicates the learning rate; This represents the output of the m-th tree; This represents the fused feature vector at time t.

[0049] In one feasible implementation, the objective function of the quantile regression tree in the gradient boosting tree model is expressed by the following formula (7): (7) in, This indicates that the objective function value of the gradient boosting tree is at the quantile level. The following losses; Indicates quantile level, This is used to control the predicted quantile locations. Indicates the time step index; Indicates the sample weight coefficient; Indicates time The true quantile values ​​of the target variable; The quantile value at time t is predicted by the gradient boosting tree model, and the quantile crossover phenomenon is eliminated by order-preserving constraints.

[0050] In one feasible implementation, after obtaining the multi-node wind power prediction results for the second stage, this embodiment of the invention introduces a post-processing strategy based on ordinal-preserving regression and quantile adaptive scaling to perform secondary optimization of the prediction interval coverage and interval width without sacrificing point prediction accuracy. The specific implementation process of this strategy includes:

[0051] Ordinal-preserving regression is used to calibrate the prediction interval probabilistically. The obtained historical wind speed sequence data and corresponding actual wind power sequence data are preprocessed, and training samples are constructed according to a preset time sliding window. These training samples are then divided into training, validation, and test sets by time or randomization. Based on the validation set data, the samples are sorted according to the initial quantiles (e.g., 5% and 95%) output by the gradient boosting tree model, and the empirical coverage rate of actual observations falling within this interval is calculated. Ordinal-preserving regression is used to fit the monotonic mapping relationship between "nominal quantiles - empirical coverage rate," thereby calibrating the quantile predictions for the test set to make the interval coverage rate closer to the preset confidence level. The gradient boosting tree model can output a single point prediction. Alternatively, quantile regression can be used to output multiple quantile predictions simultaneously. This is to characterize the probability distribution and uncertainty of the prediction results.

[0052] (2) Based on the order-preserving calibration, quantile adaptive scaling is implemented to further optimize the interval width. According to the average width and coverage performance of each interval on the validation set, the upper and lower bounds of the interval are dynamically adjusted: if the coverage is higher than the target value, the interval is shrunk proportionally; if it is lower than the target value, the interval is expanded. The scaling ratio is adaptively determined according to the prediction error distribution. The interval is appropriately expanded in periods of high uncertainty and shrunk in periods of high certainty, thereby minimizing the average width of the interval while ensuring the target coverage.

[0053] The aforementioned processing strategy is closely integrated with the two-stage prediction model proposed in this embodiment of the invention. After the gradient boosting tree model outputs the multi-quantile prediction results in the second stage, the 5% and 95% quantiles are calibrated and scaled as described above to obtain the final optimized prediction interval. This processing only introduces a small amount of linear computation in the inference stage, does not affect the model training efficiency, and is suitable for real-time wind power prediction systems.

[0054] In one feasible implementation, when the application scenario requires distribution-level prediction, this embodiment of the invention further integrates conditional distribution fitting and empirical distribution estimation methods based on scenario sets on the basis of the two-stage prediction framework to output a complete probability distribution of future wind power. This capability is built upon the multi-quantile prediction results generated by the second-stage gradient boosting decision tree. Specifically, firstly, multiple conditional quantiles (such as 5%, 50%, and 95%) synchronously output by the gradient boosting decision tree model based on the quantile loss function are used as the basic framework to characterize the prediction distribution. Secondly, a non-parametric conditional kernel density estimation method is used to smoothly fit the multiple conditional quantiles to obtain a continuous conditional probability density function, where the continuous conditional probability density function can express the possibility of wind power taking any value under given input features. To more fully characterize the spatiotemporal dependence of wind power and the joint uncertainty in multi-step prediction, embodiments of this invention can also generate a set of statistically representative future wind power scenarios based on bootstrapping sampling or Monte Carlo simulation using historical prediction errors. This set of scenarios constitutes a discrete empirical distribution of future power, which can be directly used in downstream decision-making models that require probabilistic input, such as stochastic optimal scheduling or electricity market risk assessment. The entire distribution prediction module seamlessly integrates with the core two-stage prediction model in post-processing form. During the training phase, validation data is used to calibrate the distribution parameters, and during the inference phase, only an acceptable computational overhead is added. Ultimately, it provides users with flexible and comprehensive wind power prediction information output that combines point estimation, interval estimation, and complete distribution estimation.

[0055] In one feasible implementation, regarding engineering deployment, this embodiment of the invention designs a systematic model update and structural adaptation mechanism to ensure that the two-stage prediction model maintains high accuracy and strong robustness in long-term practical operation. This mechanism comprises three core components: data-driven dynamic graph structure adaptive update, online detection and triggered retraining oriented towards concept drift, and continuous monitoring and probability calibration of model performance based on a rolling time window. The specific implementation process includes:

[0056] First, regarding the graph structure relied upon in the first stage, namely the adjacency matrix between wind turbines, this embodiment of the invention periodically re-evaluates and optimizes it. Specifically, the system periodically (e.g., weekly or monthly) recalculates the cross-correlation of wind speed sequences between wind turbines or a measure based on fluctuation similarity using the latest historical wind speed data, thereby updating the edge weights of the adjacency matrix to reflect the dynamic evolution of spatial correlations between turbines. Simultaneously, to improve computational efficiency and generalization ability, the updated adjacency matrix undergoes thresholding sparsity processing, retaining only strongly correlated connections, thus reducing the computational complexity of graph convolution while maintaining spatial information extraction capabilities.

[0057] Secondly, to address potential changes in data distribution over time, this embodiment of the invention integrates an online drift detection and trigger-based model update module. This module continuously monitors the statistical characteristics (such as mean, variance, and distribution pattern) of the input data stream (e.g., wind speed and environmental factors) as well as the model's prediction error metrics in the latest rolling window. Once a significant shift in statistical characteristics is detected or the prediction error continuously exceeds a preset threshold, the system automatically triggers a model retraining process. This process, while retaining historical knowledge, incrementally updates or partially retrains the model parameters using data from the latest time period, thereby quickly adapting to new data patterns and avoiding performance degradation due to environmental changes.

[0058] Finally, to ensure the long-term reliability of prediction results, especially probability interval predictions, this embodiment of the invention establishes a continuous performance monitoring and rolling calibration system. After deployment, the system continuously collects prediction results and actual observations, calculating various evaluation metrics for point and interval predictions within a rolling time window (e.g., the past week). Using these real-time evaluation results, the output quantiles of the second-stage gradient boosting decision tree model are calibrated online, for example, by adjusting the upper and lower bounds of the interval through ordinal regression or Bayesian methods, ensuring that the actual coverage of the prediction interval within the rolling window remains stably near the target confidence level. Through the synergistic operation of the aforementioned adaptive update, drift response, and rolling calibration mechanisms, the method proposed in this embodiment of the invention can maintain excellent prediction accuracy and stable probabilistic reliability throughout the entire lifecycle of a wind farm.

[0059] This invention presents a probabilistic prediction method for time-series data by constructing a two-stage spatiotemporal graph network and fusing multi-source information. In the first stage, a dynamically updatable adjacency matrix is ​​constructed based on the correlation of historical data fluctuations. Combined with a graph convolutional network and a spatiotemporal Transformer incorporating an external memory mechanism, this effectively characterizes the complex spatial dependencies in multi-node systems and their dynamic characteristics over time, significantly improving the accuracy and stability of intermediate variable (e.g., wind speed) predictions. The introduction of an external attention mechanism maintains global spatiotemporal dependency modeling capabilities while reducing computational and storage overhead, making this invention more suitable for high-dimensional, multi-node, and long-sequence practical engineering scenarios, and solving the problem of excessively high computational costs associated with traditional self-attention methods. Combined with the temporal feature extraction structure of a multilayer perceptron and the weighted optimization of key business intervals using a threshold-aware loss function, this invention achieves more accurate fitting and generalization in important segments affecting the final goal, effectively reducing the impact of prediction errors on decision-making.

[0060] In the second stage, this invention introduces multi-source environmental factors such as temperature, humidity, and air pressure, and employs a gradient boosting tree model to achieve deep integration of intermediate variables and external factors. This fully leverages the ensemble model's ability to characterize complex nonlinear relationships and feature interactions, improving the accuracy and interpretability of target variable (e.g., wind power) predictions. By optimizing model parameters through grid search and cross-validation, this invention achieves excellent performance across multiple accuracy and probability metrics, ensuring the prediction results possess acceptable accuracy and stability for engineering applications. Furthermore, this invention establishes an uncertainty propagation mechanism between the two stages, converting the probabilistic prediction information from the first stage into usable feature inputs for the downstream stage. This allows the second-stage model to fully perceive the prediction range and error morphology, enhancing its resilience during target mapping and reducing the cumulative amplification of upstream prediction errors.

[0061] The embodiments of this invention can significantly improve the prediction accuracy and stability of multi-node and multi-source heterogeneous time series data; it realizes the effective transmission of probabilistic information from intermediate variables to target variables, enhancing the credibility and resilience of prediction results; it optimizes the balance between coverage and width of interval prediction, providing accurate and reliable decision support for engineering scheduling; it reduces the computational overhead of large-scale spatiotemporal modeling, and has good real-time performance and deployment feasibility; the framework proposed in the embodiments of this invention has strong versatility and can be directly transferred and extended in various spatiotemporal coupled prediction scenarios such as new energy power, traffic management, weather forecasting, and industrial Internet of Things, meeting the needs of different fields for high-precision and high-reliability time series prediction.

[0062] Figure 3 This is a block diagram of a temporal probability prediction device that fuses a two-stage spatiotemporal graph network with multi-source information, provided in an embodiment of the present invention. This device is used in a temporal probability prediction method that fuses a two-stage spatiotemporal graph network with multi-source information. (Refer to...) Figure 3 The device includes an acquisition unit 310, a construction unit 320, a first prediction unit 330, and a second prediction unit 340. Wherein:

[0063] The acquisition unit 310 is used to acquire historical wind speed sequence data and corresponding real wind power sequence data from the multi-node sensor network of the wind farm short-term wind power prediction system; preprocess the historical wind speed sequence data and corresponding real wind power sequence data, and construct training samples according to a preset time sliding window. Construction unit 320 is used to construct a temporal probabilistic prediction framework for a two-stage spatiotemporal graph network and multi-source information fusion; the framework includes: a wind speed prediction model and a gradient boosting tree model; The first prediction unit 330 is used to construct an adjacency matrix between nodes based on the fluctuation correlation of historical sequence data; input the training samples and the adjacency matrix between nodes into the wind speed prediction model, train the wind speed prediction model through the threshold-aware loss function of the designed wind speed-power nonlinear relationship, obtain the trained wind speed prediction model, predict the wind speed probability of real-time input based on the trained wind speed prediction model, and output the multi-node wind speed prediction probability of the first stage. The second prediction unit 340 is used to input the multi-node wind speed prediction probability and the acquired multi-source environmental factor data into the gradient boosting tree model for the second stage prediction, and output the second stage multi-node wind power prediction result.

[0064] Optionally, the multi-source environmental factor data includes: temperature, humidity, and air pressure.

[0065] Optionally, the wind speed prediction model includes: a spatial feature extraction module and a temporal feature extraction module; The spatial feature extraction module consists of a graph convolutional network and a spatial Transformer model with an external attention mechanism. The time feature extraction module is composed of a multi-layer sensing mechanism.

[0066] Optionally, the first prediction unit 330 is configured to: The training samples and the adjacency matrix between nodes are input into the spatial feature extraction module of the wind speed prediction model. The adjacency matrix of the nodes is weighted, aggregated, and transformed by a graph convolutional network to output the temporal features of the nodes. The node temporal features and location codes are superimposed and input into a spatial Transformer model with an external attention mechanism. The external attention mechanism is used for calculation, and the weight matrix of the external attention mechanism is output. The weight matrix of the external attention mechanism is input into the time feature extraction module, and nonlinear mapping is performed through a multilayer perceptron to output the multi-node wind speed prediction probability. Based on the multi-node wind speed prediction probability, the wind speed prediction model is trained by the threshold perception loss function of the designed wind speed-power nonlinear relationship to obtain the trained wind speed prediction model. Based on the trained wind speed prediction model, the multi-node wind speed prediction probability of the first stage is predicted. The process of nonlinear mapping through a multilayer perceptron is represented by the following formula (1): (1) in, This represents the output vector at the current time step index t. Represents the MLP mapping function; Indicates from The matrix formed by concatenating the input sequence up to t according to time. Represents a non-linear activation function; This represents the weight matrix of the first layer of the MLP; This represents the weight matrix of the second layer of the MLP; This represents the bias vector of the first layer of the MLP; This represents the bias vector of the second layer of the MLP; This represents the input feature matrix obtained by concatenating the observations of the object under study in chronological order within the time window [t-w+1, t]; t represents the time step index in the prediction sequence.

[0067] Optionally, the threshold sensing loss function of the wind speed-power nonlinear relationship is expressed by the following formula (2): (2) in, This represents the total value of the weighted quantile loss function; Indicates the weighting coefficient; This represents the actual value of the current time step index t; This represents the predicted value at the current time step index t; represents the quantile level; t represents the time step index in the predicted sequence.

[0068] Optionally, the second prediction unit 340 is used for: Based on the multi-node wind speed prediction probability input pre-trained spatiotemporal network model, the quantiles corresponding to future time steps are predicted based on the pre-selected quantile level set, and the quantile prediction vector is output. The quantile prediction vector is concatenated with the features extracted from the wind speed prediction model to obtain the concatenated feature vector. The concatenated feature vector is fused with multi-source environmental factor data to obtain a fused feature vector; the fused feature vector is then input into the gradient boosting tree model for the second stage of prediction, and the second stage multi-node wind power prediction results are output.

[0069] Optionally, the process of inputting the fused feature vector into the gradient boosting tree model for the second-stage prediction is represented by the following formula (3): (3) in, This represents the predicted value of the target variable at time t; Indicates the number of weak learners; Indicates the learning rate; This represents the output of the m-th tree; The fused feature vector represents time.

[0070] This invention presents a probabilistic prediction method for time-series data by constructing a two-stage spatiotemporal graph network and fusing multi-source information. In the first stage, a dynamically updatable adjacency matrix is ​​constructed based on the correlation of historical data fluctuations. Combined with a graph convolutional network and a spatiotemporal Transformer incorporating an external memory mechanism, this effectively characterizes the complex spatial dependencies in multi-node systems and their dynamic characteristics over time, significantly improving the accuracy and stability of intermediate variable (e.g., wind speed) predictions. The introduction of an external attention mechanism maintains global spatiotemporal dependency modeling capabilities while reducing computational and storage overhead, making this invention more suitable for high-dimensional, multi-node, and long-sequence practical engineering scenarios, and solving the problem of excessively high computational costs associated with traditional self-attention methods. Combined with the temporal feature extraction structure of a multilayer perceptron and the weighted optimization of key business intervals using a threshold-aware loss function, this invention achieves more accurate fitting and generalization in important segments affecting the final goal, effectively reducing the impact of prediction errors on decision-making.

[0071] In the second stage, this invention introduces multi-source environmental factors such as temperature, humidity, and air pressure, and employs a gradient boosting tree model to achieve deep integration of intermediate variables and external factors. This fully leverages the ensemble model's ability to characterize complex nonlinear relationships and feature interactions, improving the accuracy and interpretability of target variable (e.g., wind power) predictions. By optimizing model parameters through grid search and cross-validation, this invention achieves excellent performance across multiple accuracy and probability metrics, ensuring the prediction results possess acceptable accuracy and stability for engineering applications. Furthermore, this invention establishes an uncertainty propagation mechanism between the two stages, converting the probabilistic prediction information from the first stage into usable feature inputs for the downstream stage. This allows the second-stage model to fully perceive the prediction range and error morphology, enhancing its resilience during target mapping and reducing the cumulative amplification of upstream prediction errors.

[0072] The embodiments of this invention can significantly improve the prediction accuracy and stability of multi-node and multi-source heterogeneous time series data; it realizes the effective transmission of probabilistic information from intermediate variables to target variables, enhancing the credibility and resilience of prediction results; it optimizes the balance between coverage and width of interval prediction, providing accurate and reliable decision support for engineering scheduling; it reduces the computational overhead of large-scale spatiotemporal modeling, and has good real-time performance and deployment feasibility; the framework proposed in the embodiments of this invention has strong versatility and can be directly transferred and extended in various spatiotemporal coupled prediction scenarios such as new energy power, traffic management, weather forecasting, and industrial Internet of Things, meeting the needs of different fields for high-precision and high-reliability time series prediction.

[0073] Figure 4 This is a schematic diagram of the structure of a temporal probability prediction device that integrates a two-stage spatiotemporal graph network and multi-source information fusion, as provided in an embodiment of the present invention. Figure 4As shown, the temporal probability prediction device that combines a two-stage spatiotemporal graph network with multi-source information fusion can include the above-mentioned... Figure 3 The illustrated temporal probability prediction device 410, which integrates a two-stage spatiotemporal graph network with multi-source information fusion, may optionally include a first processor 2001.

[0074] Optionally, the temporal probability prediction device 410 for the fusion of a two-stage spatiotemporal graph network and multi-source information may further include a memory 2002 and a transceiver 2003.

[0075] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0076] The following is combined with Figure 4 The components of the temporal probability prediction device 410, which integrates a two-stage spatiotemporal graph network and multi-source information fusion, are described in detail below: The first processor 2001 is the control center of the time-series probability prediction device 410 that fuses a two-stage spatiotemporal graph network with multi-source information. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0077] Optionally, the first processor 2001 can execute various functions of the time series probability prediction device 410 that integrates two-stage spatiotemporal graph network and multi-source information fusion by running or executing software programs stored in memory 2002 and calling data stored in memory 2002.

[0078] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 4 CPU0 and CPU1 are shown in the diagram.

[0079] In a specific implementation, as one example, the temporal probability prediction device 410 that fuses a two-stage spatiotemporal graph network with multi-source information can also include multiple processors, for example... Figure 4The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0080] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0081] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or exist independently, and may be connected to the interface circuit of the timing probability prediction device 410 for multi-source information fusion via a two-stage spatiotemporal graph network. Figure 4 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0082] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0083] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 4 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0084] Optionally, the transceiver 2003 can be integrated with the first processor 2001 or exist independently, and can be connected to the interface circuit of the time series probability prediction device 410, which integrates multi-source information, through a two-stage spatiotemporal graph network. Figure 4(Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0085] It should be noted that, Figure 4 The structure of the two-stage spatiotemporal graph network and multi-source information fusion temporal probability prediction device 410 shown does not constitute a limitation on the router. The actual two-stage spatiotemporal graph network and multi-source information fusion temporal probability prediction device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0086] Furthermore, the technical effect of the time series probability prediction device 410 that integrates a two-stage spatiotemporal graph network and multi-source information fusion can be referred to the technical effect of the time series probability prediction method that integrates a two-stage spatiotemporal graph network and multi-source information fusion described in the above method embodiments, and will not be repeated here.

[0087] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or it may be any conventional processor, etc.

[0088] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0089] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0090] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0091] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0092] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0093] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0095] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0096] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0097] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0098] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A temporal probability prediction method that fuses a two-stage spatiotemporal graph network with multi-source information, characterized in that, The method includes: S1. Obtain historical wind speed sequence data and corresponding real wind power sequence data from the multi-node sensor network of the short-term wind power prediction system of the wind farm; preprocess the historical wind speed sequence data and corresponding real wind power sequence data, and construct training samples according to a preset time sliding window; S2. A temporal probabilistic prediction framework for fusion of a two-stage spatiotemporal graph network and multi-source information is constructed; the framework includes: a wind speed prediction model and a gradient boosting tree model; S3. Based on the fluctuation correlation of historical sequence data, construct the adjacency matrix between nodes; input the training samples and the adjacency matrix between nodes into the wind speed prediction model, train the wind speed prediction model through the threshold perception loss function of the designed wind speed-power nonlinear relationship, obtain the trained wind speed prediction model, predict the wind speed probability of real-time input based on the trained wind speed prediction model, and output the multi-node wind speed prediction probability of the first stage. S4. Input the multi-node wind speed prediction probability and the acquired multi-source environmental factor data into the gradient boosting tree model for the second stage of prediction, and output the second stage multi-node wind power prediction results.

2. The temporal probability prediction method based on dual-stage spatiotemporal graph network and multi-source information fusion according to claim 1, characterized in that, The multi-source environmental factor data includes: temperature, humidity, and air pressure.

3. The temporal probability prediction method based on dual-stage spatiotemporal graph network and multi-source information fusion according to claim 1, characterized in that, The wind speed prediction model includes: a spatial feature extraction module and a temporal feature extraction module; The spatial feature extraction module consists of a graph convolutional network and a spatial Transformer model with an external attention mechanism. The time feature extraction module is composed of a multi-layer sensing mechanism.

4. The temporal probability prediction method based on the fusion of a two-stage spatiotemporal graph network and multi-source information as described in claim 3, characterized in that, S3 constructs an adjacency matrix between nodes based on the fluctuation correlation of historical sequence data; the training samples and the adjacency matrix between nodes are input into the wind speed prediction model, and the wind speed prediction model is trained using a threshold-aware loss function based on the designed wind speed-power nonlinear relationship to obtain a trained wind speed prediction model. Based on the trained wind speed prediction model, the real-time input wind speed probability is predicted, and the first-stage multi-node wind speed prediction probability is output, including: S31. Input the training samples and the adjacency matrix between nodes into the spatial feature extraction module in the wind speed prediction model, and perform weighted aggregation and feature transformation on the adjacency matrix of the nodes through a graph convolutional network to output the temporal features of the nodes. S32. Superimpose the node temporal features and location codes and input them into the spatial Transformer model containing the external attention mechanism. Calculate the weight distribution matrix of the external attention mechanism through the external attention mechanism. S33. Input the weight distribution matrix of the external attention mechanism into the time feature extraction module, perform nonlinear mapping through a multilayer perceptron, and output the multi-node wind speed prediction probability. Based on the multi-node wind speed prediction probability, train the wind speed prediction model through the threshold perception loss function of the designed wind speed-power nonlinear relationship to obtain the trained wind speed prediction model. Based on the trained wind speed prediction model, predict the multi-node wind speed prediction probability of the first stage. The process of nonlinear mapping through a multilayer perceptron is represented by the following formula (1): （1） in, This represents the output vector at the current time step index t. Represents the MLP mapping function; Indicates from The matrix formed by concatenating the input sequence up to t according to time. Represents a nonlinear activation function; This represents the weight matrix of the first layer of the MLP; This represents the weight matrix of the second layer of the MLP; This represents the bias vector of the first layer of the MLP; This represents the bias vector of the second layer of the MLP; This represents the input feature matrix obtained by concatenating the observations of the object under study in chronological order within the time window [t-w+1, t]; t represents the time step index in the prediction sequence.

5. The temporal probability prediction method based on dual-stage spatiotemporal graph network and multi-source information fusion according to claim 1, characterized in that, The threshold sensing loss function of the wind speed-power nonlinear relationship is expressed by the following formula (2): （2） in, This represents the total value of the weighted quantile loss function; Indicates the weighting coefficient; This represents the actual value of the current time step index t; This represents the predicted value at the current time step index t; represents the quantile level; t represents the time step index in the predicted sequence.

6. The temporal probability prediction method based on the fusion of a two-stage spatiotemporal graph network and multi-source information as described in claim 1, characterized in that, The S4 process inputs the multi-node wind speed prediction probability and the acquired multi-source environmental factor data into the gradient boosting tree model for the second-stage prediction, and outputs the second-stage multi-node wind power prediction results, including: S41. Based on the multi-node wind speed prediction probability input into the pre-trained spatiotemporal network model, the quantiles corresponding to future time steps are predicted based on the pre-selected quantile level set, and the quantile prediction vector is output. S42. Concatenate the quantile prediction vector with the features extracted from the wind speed prediction model to obtain the concatenated feature vector. S43. The spliced ​​feature vector is fused with multi-source environmental factor data to obtain a fused feature vector; the fused feature vector is input into the gradient boosting tree model for the second stage prediction, and the second stage multi-node wind power prediction result is output.

7. The temporal probability prediction method based on the fusion of a two-stage spatiotemporal graph network and multi-source information as described in claim 6, characterized in that, The process of inputting the fused feature vector into the gradient boosting tree model for the second-stage prediction is represented by the following formula (3): （3） in, This represents the predicted value of the target variable at time t; Indicates the number of weak learners; Indicates the learning rate; This represents the output of the m-th tree; The fusion feature vector represents time.

8. A temporal probability prediction device for fusing a two-stage spatiotemporal graph network with multi-source information, wherein the two-stage spatiotemporal graph network and multi-source information fusion temporal probability prediction device is used to implement the temporal probability prediction method for fusing a two-stage spatiotemporal graph network with multi-source information as described in any one of claims 1-7, characterized in that, The device includes: The acquisition unit is used to acquire historical wind speed sequence data and corresponding real wind power sequence data from the multi-node sensor network of the short-term wind power prediction system of the wind farm; preprocess the historical wind speed sequence data and corresponding real wind power sequence data, and construct training samples according to a preset time sliding window. A building block is used to construct a temporal probabilistic prediction framework that integrates a two-stage spatiotemporal graph network with multi-source information fusion; the framework includes: a wind speed prediction model and a gradient boosting tree model; The first prediction unit is used to construct an adjacency matrix between nodes based on the fluctuation correlation of historical sequence data; input the training samples and the adjacency matrix between nodes into the wind speed prediction model, train the wind speed prediction model through the threshold-aware loss function of the designed wind speed-power nonlinear relationship, obtain the trained wind speed prediction model, predict the wind speed probability of real-time input based on the trained wind speed prediction model, and output the multi-node wind speed prediction probability of the first stage. The second prediction unit is used to input the multi-node wind speed prediction probability and the acquired multi-source environmental factor data into the gradient boosting tree model for the second stage prediction, and output the second stage multi-node wind power prediction result.

9. A temporal probability prediction device that fuses a two-stage spatiotemporal graph network with multi-source information, characterized in that, The time-series probability prediction device that integrates the two-stage spatiotemporal graph network with multi-source information includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

Images