Multi-time scale charging load and energy storage collaborative probabilistic forecasting method

Abstract

The application discloses a multi-time-scale charging load and energy storage collaborative probability prediction method, relates to the technical field of power systems and smart power grids, and comprises the following steps: obtaining historical multi-source data of a target system and constructing causal subgraphs of different time scales; coupling the causal subgraphs into a time-cause feedback graph atlas through a structural mapping function, wherein the atlas contains explicit time-scale control nodes; performing cross-time-scale causal effect calculation based on the time-cause feedback graph atlas to generate a causal effect matrix; constructing a joint probability prediction model by using the causal effect matrix to output a first prediction result; and triggering a time-scale switching mechanism to generate a second prediction result when it is monitored that the deviation of the first prediction result exceeds a dynamic threshold. The application realizes multi-scale causal modeling by constructing a time-cause feedback graph atlas containing time-scale control nodes, combines a dynamic scale switching mechanism and a residual compensation mechanism, significantly improves prediction accuracy and scheduling adaptability, and realizes multi-time-scale collaborative optimization management.

Description

Technical Field

[0001] This invention relates to the field of power system and smart grid technology, and more specifically, to a method for probabilistic prediction of multi-timescale charging load and energy storage coordination. Background Technology

[0002] With the rapid growth of distributed power sources and charging loads in power systems, grid loads exhibit high volatility and multi-timescale characteristics. Charging loads, including electric vehicles and energy storage facilities, show significant variations in power demand at the minute, hour, and even day levels. Energy storage systems, through energy storage and release, regulate grid balance and improve system stability at different timescales. Therefore, multi-timescale charging load and energy storage-based collaborative prediction has become an important research direction for smart grid load management and scheduling optimization.

[0003] In existing technologies, the prediction of load and energy storage across multiple time scales mainly relies on statistical models, time series analysis, or machine learning methods. Typical approaches include short-term load regression based on historical data, long-term load trend analysis, and prediction of energy storage charging and discharging using neural networks or random forest models. However, these methods typically struggle to simultaneously capture the causal relationships and interactions between different time scales, resulting in limited prediction accuracy and a lack of dynamic adjustment mechanisms to cope with rapid changes in load and energy storage behavior. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a multi-timescale charging load and energy storage collaborative probability prediction method. By constructing a time-dependent feedback graph containing time-scale control nodes, the method realizes the structural mapping and coupling of causal subgraphs at different time scales, improves the accuracy of load prediction and the adaptability of scheduling, and achieves deep fusion and refined collaborative management of multi-timescale information.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A multi-timescale probabilistic prediction method for charging load and energy storage coordination includes the following steps: acquiring historical multi-source data of the target system and constructing causal subgraphs at different time scales; coupling the causal subgraphs into a time-dependent feedback graph using a structure mapping function, wherein the graph contains explicit time-scale control nodes; performing cross-time-scale causal effect calculations based on the time-dependent feedback graph to generate a causal effect matrix; constructing a joint probabilistic prediction model using the causal effect matrix and outputting a first prediction result; and triggering a time-scale switching mechanism to generate a second prediction result when the deviation of the first prediction result exceeds a dynamic threshold.

[0007] In a preferred embodiment, the step of acquiring historical multi-source data of the target system and constructing causal subgraphs at different time scales specifically involves: acquiring historical multi-source data, performing time alignment using a dynamic time warping algorithm to generate a synchronous time series dataset; decomposing the dataset using a hybrid algorithm of wavelet packet decomposition and empirical mode decomposition to obtain subsets including short-term, medium-term, and long-term time scales; constructing an initial causal subgraph based on the subsets using a combination of conditional independence test and correlation analysis; and performing structural optimization of the initial causal subgraph based on statistical significance, removing causal edges with significance below a preset threshold to generate the final causal subgraph.

[0008] In a preferred embodiment, the coupling of the causal subgraph into a temporal feedback graph via a structural mapping function specifically involves: extracting node and edge information from short-term, medium-term, and long-term causal subgraphs to construct a causal variable space matrix of a unified dimension; calculating the structural similarity and causal consistency of semantically identical nodes across multiple time scales based on the matrix to generate a node similarity matrix; establishing a cross-time scale node mapping index based on the node similarity matrix; connecting the mapping nodes through cross-scale causal edges to form a coupled causal graph based on the node mapping index and according to predefined cross-scale coupling rules; introducing time scale control nodes into the coupled causal graph and constructing cross-scale feedback paths using these time scale control nodes as intermediaries to generate a temporal feedback graph.

[0009] In a preferred embodiment, the method for selecting the time-scale control nodes is as follows: candidate time-scale control nodes are selected based on the semantic and functional similarity of key feature variables in the multi-time-scale causal subgraph; a comprehensive control index of the candidate nodes is calculated through a weighted evaluation model, wherein the model integrates three types of indicators: causal strength, node frequency characteristics, and prediction sensitivity; and nodes whose comprehensive control index exceeds a dynamic threshold are selected to form the final time-scale control node set.

[0010] In a preferred embodiment, the step of performing cross-timescale causal effect calculation based on the time-based feedback graph to generate a causal effect matrix specifically involves: extracting key feature nodes and their causal paths at multiple time scales from the time-based feedback graph to generate a set of causal paths; calculating the initial causal effect of each path on the target variable using a causal inference algorithm to form an initial causal effect matrix; correcting the causal weights of the paths and adjusting the confidence intervals based on the pointing relationships of the time-scale control nodes to generate a final causal effect matrix; and incrementally correcting the time-based feedback graph based on the final causal effect matrix through a feedback mechanism.

[0011] In a preferred embodiment, the incremental correction of the time-cause feedback graph based on the final causal effect matrix through a feedback mechanism specifically involves: calculating the rate of change of the causal contribution of key feature variables based on the final causal effect matrix; marking variables as structural drift variables when the rate of change exceeds a preset rate of change threshold; triggering a local causal structure incremental correction mechanism to adjust the topological relationships of nodes and edges related to drift variables; applying the minimum description length criterion to select the optimal solution from candidate adjustment schemes and outputting the corrected time-cause feedback graph.

[0012] In a preferred embodiment, the step of constructing a joint probabilistic prediction model using the causal effect matrix and outputting a first prediction result specifically involves: decomposing the final causal effect matrix to separate short-term, medium-term, and long-term direct causal effect sub-matrices and cross-scale indirect causal effect sub-matrices; constructing corresponding probabilistic prediction sub-models based on the direct causal effect sub-matrices of each time scale; establishing feature projection channels using the cross-scale indirect causal effect sub-matrices to achieve stepwise feature mapping from short-term to medium-term to long-term; backpropagating the long-term prediction result to the medium-term and short-term sub-models for correction through feedback connections; and fusing the outputs of the sub-models at each scale using the Bayesian model averaging method to generate the first prediction result.

[0013] In a preferred embodiment, the step of triggering a time scale switching mechanism when the deviation of the first prediction result exceeds a dynamic threshold specifically involves: comparing the first prediction result with real-time power grid load monitoring data to calculate the prediction deviation at each time scale; triggering scale switching when the deviation at any scale exceeds a preset threshold or when multiple scale prediction results conflict; constructing a three-dimensional evaluation vector that includes the deviation change rate trend, causal stability coefficient, and scheduling priority weight; calculating the score at each scale using a preset evaluation function, selecting the scale with the highest score as the optimal prediction scale; and outputting a time scale switching command to the prediction system.

[0014] In a preferred embodiment, generating the second prediction result specifically involves: selecting the model corresponding to the optimal time scale as the main prediction model based on the time scale switching instruction, and setting the models at other scales as auxiliary models; calculating the fusion weight coefficients of the main and auxiliary models based on the scheduling priority weights in the three-dimensional evaluation vector; backpropagating the prediction bias to the auxiliary model through the cross-scale causal path in the time-cause feedback graph; injecting a residual compensation tensor into the representation layer of the auxiliary model, calculating the residual gradient and slightly updating the weights; and fusion of the main model output and the auxiliary model output according to the weight coefficients to generate the second prediction result.

[0015] In a preferred embodiment, the step of backpropagating the prediction bias to the auxiliary model through the cross-scale causal path in the time-cause feedback graph specifically involves: identifying key feature variables based on the prediction bias and retrieving the set of cross-scale causal paths associated with the key feature variables; calculating the sensitivity adjustment factor of each feature variable in the auxiliary model based on the causal path set; performing perturbation simulation on the input features of each auxiliary model and calculating the potential residual contribution of each variable by weighting it with the sensitivity adjustment factor; and constructing a residual compensation tensor based on the potential residual contribution, perturbation amplitude, and causal strength.

[0016] The technical effects and advantages of the multi-timescale charging load and energy storage synergistic probabilistic prediction method of this invention are as follows:

[0017] 1. This invention constructs a multi-timescale causal subgraph and couples it into a time-dependent feedback graph, which can clearly characterize the direct causal effects and cross-scale indirect causal effects between short-term, medium-term and long-term charging loads and energy storage systems. It realizes the step-by-step mapping and correction of prediction results at each time scale, thereby significantly improving the accuracy of load and energy storage co-prediction and reducing prediction deviations caused by data noise or local fluctuations.

[0018] 2. This invention achieves prediction optimization through a joint prediction and dynamic switching mechanism driven by a causal effect matrix, combined with Bayesian model fusion and residual compensation strategies. This mechanism breaks through the static mode of traditional multi-model weighted fusion, realizing dynamic weight allocation and scale adaptive switching based on causal strength, providing a refined error correction path for multi-timescale prediction. This dynamic optimization not only enhances the robustness of the prediction results but also expands the solution space of the prediction model through cross-scale feedback correction, improving prediction accuracy in complex scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the multi-timescale charging load and energy storage synergistic probability prediction method of the present invention.

[0020] Figure 2 This is a schematic diagram of the second prediction result process of the present invention. Detailed Implementation

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

[0022] Example 1, Figure 1 This invention presents a probabilistic prediction method for multi-timescale charging load and energy storage coordination, comprising:

[0023] S1, Obtain historical multi-source data of the target system and construct causal subgraphs at different time scales;

[0024] In this implementation case, historical multi-source data of the target system is obtained, and causal subgraphs at different time scales are constructed, as follows:

[0025] Historical multi-source data was acquired through user-end apps, power companies, and meteorological service platforms. This data included, but was not limited to, charging load data from electric vehicle charging stations, charging and discharging status data from energy storage systems, power grid operating parameters, user charging behavior records, and weather and holiday data. Since these data sources had different sampling frequencies and recording periods, time alignment processing was first required to ensure temporal consistency in subsequent causal analysis.

[0026] A dynamic time warping algorithm is employed to align heterogeneous data from multiple sources, minimizing nonlinear offsets on the time axis. Based on the global shortest distance criterion, this algorithm identifies the optimal alignment path between different time series and maps data interpolated at different times to a unified reference time axis, thereby obtaining a time series dataset with consistent structure and strong synchronization, which serves as the input basis for causal modeling.

[0027] After the time series data is synchronized, a hybrid algorithm combining wavelet packet decomposition (WPD) and empirical mode decomposition (EMD) is used to scale the dataset. WPD captures high-frequency instantaneous features in the signal, while EMD adaptively extracts nonlinear trend components. This hybrid decomposition strategy effectively extracts subsets at three time scales: short-term (daily), medium-term (weekly), and long-term (monthly). For each subset at a given time scale, a causal structure learning method based on conditional independence testing and statistical correlation analysis is used to construct the corresponding causal subgraph. Specifically, an improved PC algorithm is used to perform conditional independence testing, identifying the independent relationships between variables under given conditions. Combined with statistical correlation measures (such as partial correlation coefficients and mutual information), the significance and strength of edges are evaluated to generate initial causal subgraphs for each time scale.

[0028] To ensure the statistical validity and engineering usability of the initial causal subgraph, a structural verification of the initial causal graph is further performed. A saliency screening mechanism is introduced to remove weakly correlated edges with weights below a set confidence threshold, thus preventing noisy paths from interfering with the inference results. The final causal subgraphs at each time scale, after saliency screening, are output as the basic input for subsequent graph coupling and causal inference.

[0029] S2, the causal subgraph is coupled into a time-dependent feedback graph through a structural mapping function, the graph containing explicit time-scale control nodes;

[0030] In this implementation case, the causal subgraph is coupled into a time-cause feedback graph through a structure mapping function, specifically as follows:

[0031] First, node and edge information is extracted from the short-term, medium-term, and long-term causal subgraphs, respectively. Each causal subgraph is defined as a directed graph:

[0032] ,

[0033] in, For the set of causal variable nodes at the corresponding time scale, Let be the set of causal edges between nodes.

[0034] To achieve a unified representation of causal variables across different time scales, a causal variable representation space is designed, mapping variables from all time scales to a common feature space. This space uses the semantic features, statistical features, and physical attributes of nodes to perform vectorized encoding, forming a feature vector for each node. ,in As a feature dimension, the feature vectors of all nodes at the same time scale are arranged row-wise to form a causal variable space matrix. :

[0035]

[0036] in, This represents the total number of nodes.

[0037] Based on the spatial matrix of causal variables at various time scales, the structural similarity and causal consistency between nodes with the same semantics or similar functions are calculated for subsequent node mapping.

[0038] Structural similarity is calculated using cosine similarity:

[0039] , ,

[0040] in, and These represent the short-term and medium-term time scales, respectively. and These are the node feature vectors corresponding to the short-term and medium-term time scales, respectively. Causal consistency is measured by comparing the similarity of causal paths between nodes, and is defined as the similarity of the adjacency matrices of nodes in their respective causal subgraphs, using the Jaccard similarity coefficient or path consistency index.

[0041]

[0042] in, Representing the adjacency set of nodes, and combining structural similarity and causal consistency, a node similarity matrix is ​​formed:

[0043]

[0044] in, The hyperparameters are used to adjust the weights of both.

[0045] Based on the node similarity matrix Using a maximum matching algorithm or a threshold filtering strategy, the one-to-one mapping relationship between nodes across time scales is determined, generating a set of node mapping indexes.

[0046]

[0047] in, This is a similarity threshold used to remove low-similarity mappings and ensure mapping quality.

[0048] Based on node mapping index According to predefined cross-scale coupling rules, nodes at different time scales are connected by cross-scale causal edges to form a coupled causal graph. ,in, , , It is a cross-scale causal edge set.

[0049] The method for selecting the time-scale control nodes is as follows:

[0050] In the process of constructing the causal feedback graph, in order to achieve effective coordination and information transmission of causal paths between different time scales, it is necessary to introduce time scale control nodes into the coupled causal graph as intermediary hubs for feedback relationships between scales.

[0051] From the constructed short-term, medium-term, and long-term causal subgraphs, key feature variable nodes are extracted. These nodes, combined with their semantic labels, functional types, and predictive relevance indicators, serve as the data basis for time-scale control node selection. All data must first be processed in a standardized format, and the feature variable nodes must be standardized and coded to facilitate cross-scale comparisons.

[0052] The semantic similarity between node pairs is calculated and the consistency of functional types is judged. The variable type to which the node belongs (such as load, behavior, environment) is matched with labels. Nodes with inconsistent types are eliminated, and the set of nodes that meet the threshold requirements in both semantics and function is selected as the candidate time scale control node set.

[0053] For the set of candidate timescale control nodes Each node in Calculate the following indicators: causal strength Node frequency characteristics and prediction sensitivity .

[0054] Causal strength This indicates the degree of causal influence between a node and the target predictor variable, and its calculation formula is:

[0055]

[0056] in, For time scale sets, Indicated in scale Next node Pointing to target variable The edge weights.

[0057] Node frequency characteristics The frequency of a node in the causal graph at each time scale is represented by the following formula:

[0058]

[0059] in, A certain time scale The causal subgraph below.

[0060] Predictive sensitivity The degree to which small perturbations in the node input values ​​affect the prediction results can be calculated through partial derivative sensitivity analysis:

[0061]

[0062] in, For the output of the prediction model, This is the input node.

[0063] Causality strength Node frequency characteristics and prediction sensitivity After normalizing the three indicators, calculate the comprehensive score for each candidate node:

[0064]

[0065] in, , , These are the normalized results for the corresponding indicators. , , Weight parameters, satisfying It can be adjusted according to the prediction task (e.g., when prioritizing prediction stability). (Can be set to a larger value). A preset comprehensive score threshold is used; nodes with comprehensive scores greater than this threshold are selected as the final time scale control node set. .

[0066] Cross-scale feedback edges are constructed using time-scale control nodes as intermediaries, connecting them with causal nodes at various scales to form feedback paths. Examples include "short-term charging load → control node (daily total charging) → medium-term temperature" and "long-term charging trend → control node (monthly average temperature) → short-term temperature." This enables bidirectional flow and dynamic adjustment of causal information across different scales, improving the accuracy and stability of the prediction model. The final output is a temporal-causal feedback graph containing time-scale control nodes and cross-scale feedback paths. ,in , , Cross-scale feedback edge set.

[0067] S3, based on the time-cause feedback map, performs causal effect calculations across time scales and generates a causal effect matrix;

[0068] In this implementation case, causal effect calculations across time scales are performed based on the time-cause feedback graph to generate a causal effect matrix, specifically:

[0069] From the time-factor feedback graph Key feature nodes and their causal paths at different time scales are extracted. Key feature nodes refer to nodes that have a significant impact on target variables (such as charging load and energy storage status), and can be screened by indicators such as node degree (number of connected edges) and causal strength. For example, "real-time charging power" and "instantaneous temperature" in the short-term scale, "total daily charging" and "average daily temperature" in the medium-term scale, "total monthly charging" and "average monthly temperature" in the long-term scale, and time scale control nodes such as "average daily temperature" and "total weekly charging" are all selected as key feature nodes.

[0070] A causal path is a directed path from a key feature node to the target variable, and each path consists of a series of connected causal edges. By traversing the entire temporal feedback graph, all paths from key feature nodes to the target variable are collected, and the set of causal paths is output. ,in, Indicates the first A causal path.

[0071] According to the set of causal paths The causal effect of feature variables on the target variable on each path is calculated using causal inference algorithms (such as Do-calculus). For each causal path... ,in, As key feature variables, Let be the target variable, and its causal effect be the product of the weights of each edge on the path, as shown in the formula:

[0072]

[0073] in, For the first on the path The weight of the edge, For the same feature variable To target variable The total causal effect is the sum of the causal effects of each path.

[0074] Based on the above calculations, output the initial causal effect matrix. ,in, The number of key feature variables, The number of target variables, matrix elements Indicates the first The key feature variable for the first The initial causal effect size of each target variable.

[0075] The process of controlling the pointing relationship of nodes based on the time scale, correcting the path causal weights and adjusting the confidence intervals to generate the final causal effect matrix is ​​as follows:

[0076] The time-scale control node has a regulatory effect on the node it points to, for a certain causal path. If the path contains nodes controlled by the time scale The node it points to Then the causal weight of that path needs to be multiplied by the control coefficient. The revised formula for path causality is:

[0077]

[0078] Simultaneously, the confidence interval for the causal effect is adjusted based on the influence of the control nodes. Let the initial confidence interval be:

[0079]

[0080] in, The standard deviation of the causal effect.

[0081] The corrected confidence interval is:

[0082]

[0083] in, .

[0084] Through the above corrections, the final causal effect matrix is ​​obtained. .

[0085] The incremental correction of the time-cause feedback map based on the final causal effect matrix through a feedback mechanism is specifically as follows:

[0086] Based on the final causality matrix Calculate the rate of change of causal contribution for each key characteristic variable. The rate of change of causal contribution refers to the relative change of the current causal effect compared to the causal effect at the previous time point, and the formula is:

[0087]

[0088] in, For the previous moment The key feature variable for the first The final causal effect of the target variable, For the first The rate of change of the causal contribution of the key characteristic variables.

[0089] Set a preset rate of change threshold If the rate of change of the causal contribution of any key characteristic variable If a variable is identified as a structural drift variable, a local causal structure incremental correction mechanism is triggered based on it, performing topological adjustments on relevant nodes and their edges in the time-cause feedback graph. Relevant nodes include the structural drift variable itself, nodes directly connected to it, and nodes indirectly related through causal paths. Adjustment methods include adjusting edge weights and adding or deleting edges. For edges related to the structural drift variable, the weights are adjusted according to the magnitude of the causal effect, using the following formula:

[0090]

[0091] Among them, the adjustment coefficient Control the magnitude of weight adjustment. If the causal effect between the structural drift variable and a certain node is significantly enhanced (exceeding the threshold for adding an edge, such as 1.5 times the original effect), then add an edge; if the causal effect is significantly weakened (below the threshold for deleting an edge, such as 0.5 times the original effect), then delete the edge.

[0092] The optimal tuning scheme is selected using the Minimum Description Length (MDL) criterion. The MDL criterion considers both model complexity and the goodness of fit to the data, and its formula is:

[0093]

[0094] in, The length of the structure encoding of the feedback graph after adjustment (such as the encoding length of the number of nodes, the number of edges, etc.). The encoding length for the data's fitting error based on the adjusted map is determined (a smaller fitting error results in a shorter encoding length). The adjustment scheme with the smallest MDL value is selected to generate the corrected time-cause feedback map. .

[0095] S4. Construct a joint probability prediction model using the causal effect matrix and output the first prediction result;

[0096] In this implementation case, a joint probability prediction model is constructed using a causal effect matrix, and the first prediction result is output as follows:

[0097] The final causal effect matrix is ​​decomposed to separate short-term, medium-term, and long-term direct causal effect sub-matrices, as well as cross-scale indirect causal effect sub-matrices. The direct causal effect sub-matrices reflect the causal effects of key feature variables on the target variable within the same time scale. Elements corresponding to the key feature variables and the target variable within the same time scale in the final causal effect matrix are extracted to construct the direct causal effect sub-matrices, including the short-term direct causal effect sub-matrice. Mid-term direct causal effect sub-matrix and long-term direct causal effect submatrix .

[0098] The cross-scale indirect causal effects submatrix reflects the causal effects of key feature variables on the target variable across different time scales, including the short- to medium-term cross-scale indirect causal effects submatrix. and the submatrix of indirect causal effects across medium to long-term scales .

[0099] A short-term probabilistic prediction sub-model is constructed using an LSTM (Long Short-Term Memory) network, taking short-term key feature variables as input and combining them with a short-term direct causal effect sub-matrix. The model uses an LSTM network to capture short-term dependencies in time series data and outputs the probability distribution of the short-term target variable. The output layer employs a softmax activation function to obtain the probability distribution. ,in, For short-term input features, It is a short-term target variable.

[0100] A GRU (Gated Recurrent Unit) was selected to construct a mid-term probabilistic prediction sub-model. This model utilizes key mid-term feature variables and a mid-term direct causal effect sub-matrix. GRU networks can effectively handle medium-length time series and output the probability distribution of the target variable in the medium term. ,in, For intermediate input features, This is the target variable for the medium term.

[0101] A long-term probability prediction sub-model was constructed using ARIMA (Autoregressive Integrated Moving Average). This model combines long-term key feature variables and a long-term direct causal effect sub-matrix. The ARIMA model outputs the probability distribution of the long-term target variable through operations such as stationarization, autoregression, and moving average on the time series. ,in, For long-term input features, It is the long-term target variable.

[0102] Feature projection channels are established using a cross-scale indirect causality submatrix to achieve stepwise feature mapping from short-term to medium-term to long-term. This is based on a cross-scale indirect causality submatrix from short-term to medium-term. By establishing a feature projection channel through a fully connected layer, the output features of the short-term sub-model are mapped to the medium-term scale. The mapping formula is as follows:

[0103]

[0104] in, This is the output feature vector of the short-term sub-model. The weight matrix (its elements are...) (related) For bias terms, It is the ReLU activation function. This is the feature vector mapped to the intermediate scale.

[0105] Similarly, based on the mid- to long-term cross-scale indirect causal effect submatrix Establish a feature projection channel to map the output features of the intermediate sub-model to the long-term scale. The mapping formula is as follows:

[0106]

[0107] in, This is the output feature vector of the intermediate sub-model. This is the weight matrix. For bias terms, This is the feature vector mapped to a long-term scale.

[0108] The long-term prediction results are backpropagated to the medium- and short-term sub-models via feedback connections to perform cross-scale prediction corrections. The prediction results output by the long-term sub-model are then... The feedback is passed to the intermediate sub-model to correct its predictions. The correction formula is as follows:

[0109]

[0110] in, These are the original prediction results from the intermediate sub-model. The correction coefficient (determined based on cross-scale causal effects) is used. This represents the prediction results of the intermediate sub-model based on long-term features.

[0111] Similarly, the corrected interim forecast results The results are then fed back to the short-term sub-model via a feedback channel to correct the short-term predictions.

[0112] The outputs of the sub-models at each scale are fused using the Bayesian model averaging method to obtain the first prediction result. The Bayesian model averaging method assigns different weights to each sub-model based on their prediction performance. The weight calculation formula is as follows:

[0113]

[0114] in, For the first The negative log-likelihood value of each sub-model (reflecting the goodness of fit of the model; the smaller the value, the better the fit). For the first The weights of each sub-model It is the sum of the negative log-likelihood exponents of each sub-model. These correspond to short-term, medium-term, and long-term sub-models, respectively. For the first The negative log-likelihood value of a sub-model. Specifically, if the negative log-likelihood value of a sub-model is... The smaller the value (i.e., the better the model fit). A larger value indicates a higher proportion in the total, corresponding to the weight of the sub-model. The weights of sub-models with better fits are larger; conversely, the weights of sub-models with poorer fits are smaller. In this way, the weights can objectively reflect the predictive performance of each sub-model, making the fusion result more dependent on the better-performing model.

[0115] The probability distribution of the first prediction result after fusion is as follows:

[0116]

[0117] in, , , These are the probability distributions of the corrected short-term, medium-term, and long-term sub-models, respectively. The target variable after fusion.

[0118] The first prediction result is obtained through the above fusion. .

[0119] S5, when the deviation of the first prediction result is detected to exceed the dynamic threshold, the time scale switching mechanism is triggered to generate a second prediction result;

[0120] In this implementation case, the time scale switching mechanism is triggered as follows:

[0121] The initial forecast result is compared with the real-time monitored power grid load data to calculate the forecast deviation. The forecast deviation is measured using the root mean square error (RMSE), and the formula is as follows:

[0122]

[0123] in, The first prediction result The predicted value at each moment, The first in the real-time monitored power grid load data The actual value at each moment. The number of data samples is specified. For different time scales (short-term, medium-term, and long-term), the prediction deviation is calculated for each time scale. The prediction deviation for each time scale is compared with a preset deviation threshold. If the prediction deviation for any time scale is higher than its corresponding preset deviation threshold, or if there is a conflict between the prediction results for multiple time scales (e.g., short-term predicted load increases, while medium-term and long-term predicted load decreases without reasonable basis), a time scale switch is triggered.

[0124] Construct a three-dimensional evaluation vector that includes the trend of deviation change rate, causal stability coefficient, and scheduling priority weight. Trend of deviation change rate The formula reflects the direction and rate of change of prediction deviation, and is as follows:

[0125]

[0126] in, The prediction deviation at the current moment. This represents the prediction bias from the previous time step. Causal stability coefficient. The stability of causal relationships in the time-feedback graph is measured based on the causal effect matrix, using the following formula:

[0127]

[0128] in, The standard deviation of the final causal effect matrix. The mean of the final causal effect matrix. The closer a value is to 1, the more stable the causal relationship. (Schedule priority weight) The settings are based on the grid dispatching requirements and reflect the importance of forecast results at different time scales in dispatching.

[0129] Based on the three-dimensional evaluation vector, scores for each scale are calculated using a pre-defined evaluation function. The evaluation function employs a weighted summation method, and the formula is as follows:

[0130]

[0131] in, For the first Each time scale ( The score of ) , , These are the weighting coefficients. , , , The first The trend of the rate of change of deviation for each time scale, the causal stability coefficient, and the scheduling priority weight are calculated. The scale corresponding to the highest score is selected as the optimal prediction scale, and a time scale switching instruction is output. For example, if the short-term scale score is the highest, the switching instruction is to use the short-term scale as the main prediction scale.

[0132] The second prediction result is obtained as follows:

[0133] Based on the time scale switching command, the model corresponding to the optimal time scale is selected as the primary prediction model, and models at other scales are designated as auxiliary models. For example, if the switching command is for the short-term scale, then the short-term probability prediction sub-model is the primary prediction model, and the medium-term and long-term probability prediction sub-models are auxiliary models. The fusion weight coefficients of the primary and auxiliary models are calculated based on the scheduling priority weights in the three-dimensional evaluation vector. Let the scheduling priority weight corresponding to the primary model be... The scheduling priority weights corresponding to the auxiliary model are: , The formula for calculating the fusion weight coefficient is:

[0134]

[0135]

[0136] And satisfy .

[0137] Identify key feature variables (such as "temperature" and "traffic flow" that cause significant bias) based on prediction bias, and retrieve a set of cross-scale causal paths associated with these key feature variables. Based on cross-scale causal path sets Calculate the sensitivity adjustment factor for each feature variable within the auxiliary model. , The larger the value, the higher the sensitivity of the feature variable to prediction bias. Perturbation simulations are performed on the input features of each auxiliary model (e.g., increasing or decreasing the feature value by 5%), and the simulation results are weighted using a sensitivity adjustment factor to calculate the potential residual contribution of each feature variable in the current prediction scenario:

[0138]

[0139] in, For characteristic variables The amount of change in the prediction result after the perturbation.

[0140] Based on potential residual contribution The corresponding disturbance amplitude (e.g., 5% eigenvalues) and causal strength Construct residual compensation tensor The residual compensation tensor is injected into the intermediate representation layer of the auxiliary model. The residual gradient between the intermediate representation layer and the output is calculated. The weights of the intermediate layer are then slightly updated based on the residual gradient (learning rate is set to 0.01), and the adjusted auxiliary prediction result is output. .

[0141] Prediction results of the main model based on fusion weight coefficients and auxiliary prediction results Perform weighted fusion to generate a second prediction result:

[0142]

[0143] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0144] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0145] Those skilled in the art will recognize that the modules 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 implementation should not be considered beyond the scope of this application.

[0146] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0147] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0148] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-timescale probabilistic prediction method for charging load and energy storage coordination, characterized in that, Includes the following steps: Obtain historical multi-source data of the target system and construct causal subgraphs at different time scales; The causal subgraph is coupled into a time-dependent feedback graph using a structural mapping function. Specifically, this involves: establishing a cross-timescale node mapping index based on the node similarity matrix; connecting the mapping nodes with cross-scale causal edges to form a coupled causal graph based on the node mapping index and according to predefined cross-scale coupling rules; introducing timescale control nodes into the coupled causal graph and constructing cross-scale feedback paths using these timescale control nodes as intermediaries to generate the time-dependent feedback graph; the graph contains explicit timescale control nodes. Based on the time-dependent feedback graph, perform causal effect calculations across time scales to generate a causal effect matrix; A joint probabilistic prediction model is constructed using the causal effect matrix to output the first prediction result. Specifically, the final causal effect matrix is ​​decomposed to separate the short-term, medium-term, and long-term direct causal effect sub-matrices and the cross-scale indirect causal effect sub-matrices. Based on the direct causal effect sub-matrices of each time scale, the corresponding probabilistic prediction sub-models are constructed and the prediction sub-models are corrected. The outputs of the sub-models at each scale are fused using the Bayesian model averaging method to generate the first prediction result. When the deviation of the first prediction result exceeds the dynamic threshold, the time scale switching mechanism is triggered to generate a second prediction result.

2. The multi-timescale charging load and energy storage synergistic probability prediction method according to claim 1, characterized in that, The acquisition of historical multi-source data of the target system and the construction of causal subgraphs at different time scales specifically involve: Historical multi-source data is acquired, and time alignment is performed using a dynamic time warping algorithm to generate a synchronized time series dataset; A hybrid algorithm combining wavelet packet decomposition and empirical mode decomposition is used to decompose the dataset, resulting in subsets with short-term, medium-term, and long-term time scales. Based on the aforementioned subset of data, an initial causal subgraph is constructed using a combination of conditional independence tests and correlation analysis. The initial causal subgraph is structurally optimized based on statistical significance, and causal edges with significance below a preset threshold are removed to generate the final causal subgraph.

3. The multi-timescale charging load and energy storage synergistic probability prediction method according to claim 2, characterized in that, The node similarity matrix is ​​obtained using the following method: Extract node and edge information from short-term, medium-term, and long-term causal subgraphs to construct a causal variable space matrix of a unified dimension; Based on the matrix, the structural similarity and causal consistency of semantically identical nodes across multiple time scales are calculated, generating a node similarity matrix.

4. The multi-timescale charging load and energy storage synergistic probability prediction method according to claim 3, characterized in that, The method for selecting the time-scale control nodes is as follows: Candidate time-scale control nodes are selected based on the semantic and functional similarity of key feature variables in multi-time-scale causal subgraphs. The comprehensive control index of candidate nodes is calculated by a weighted evaluation model, which integrates three types of indicators: causal strength, node frequency characteristics, and prediction sensitivity. The final time-scale control node set is composed of nodes whose comprehensive control index exceeds the dynamic threshold.

5. The multi-timescale charging load and energy storage synergistic probability prediction method according to claim 4, characterized in that, The calculation of causal effects across time scales based on the time-factor feedback map, generating a causal effect matrix, specifically involves: Extract key feature nodes and their causal paths across multiple time scales from the time-feedback graph to generate a set of causal paths. The initial causal effect of each path on the target variable is calculated using a causal reasoning algorithm, forming an initial causal effect matrix; Based on the pointing relationship of the control nodes on the time scale, the path causal weights are corrected and the confidence intervals are adjusted to generate the final causal effect matrix; Based on the final causal effect matrix, the time-cause feedback map is incrementally corrected through a feedback mechanism.

6. The multi-timescale charging load and energy storage synergistic probability prediction method according to claim 5, characterized in that, The incremental correction of the time-cause feedback map based on the final causal effect matrix through a feedback mechanism is specifically as follows: Based on the final causal effect matrix, calculate the rate of change of the causal contribution of key feature variables; When the rate of change exceeds a preset rate of change threshold, it is marked as a structural drift variable; Trigger the local causal structure incremental correction mechanism to adjust the topological relationships of nodes and edges related to drift variables; The minimum description length criterion is applied to select the optimal solution from the candidate adjustment schemes, and the corrected time-cause feedback map is output.

7. The multi-timescale charging load and energy storage synergistic probabilistic prediction method according to claim 6, characterized in that, The correction of the prediction sub-model specifically involves: By utilizing the cross-scale indirect causal effect submatrix, feature projection channels are established to achieve stepwise feature mapping from short-term to medium-term to long-term. The long-term forecast results are backpropagated to the medium- and short-term sub-models for correction via feedback connections.

8. The multi-timescale charging load and energy storage synergistic probability prediction method according to claim 7, characterized in that, When the deviation of the first prediction result is detected to exceed the dynamic threshold, a time scale switching mechanism is triggered, specifically as follows: The first prediction result is compared with the real-time power grid load monitoring data, and the prediction deviation at each time scale is calculated. When any scale deviation exceeds a preset threshold or when multi-scale prediction results conflict, scale switching is triggered. Construct a three-dimensional evaluation vector that includes the trend of deviation change rate, causal stability coefficient, and scheduling priority weight; The scores of each scale are calculated by a preset evaluation function, and the scale with the highest score is selected as the optimal prediction scale. Output time scale switching instructions to the prediction system.

9. The multi-timescale charging load and energy storage synergistic probabilistic prediction method according to claim 8, characterized in that, The generation of the second prediction result specifically involves: Based on the time scale switching instruction, the model corresponding to the optimal time scale is selected as the main prediction model, and the models at other scales are set as auxiliary models. The fusion weight coefficient of the primary and secondary models is calculated based on the scheduling priority weight in the three-dimensional evaluation vector. The prediction bias is backpropagated to the auxiliary model through the cross-scale causal path in the time-factor feedback graph; Inject residual compensation tensors into the representation layer of the auxiliary model, calculate the residual gradient, and update the weights slightly. The outputs of the main model and the auxiliary model are weighted and fused according to the weight coefficients to generate a second prediction result.

10. The multi-timescale charging load and energy storage synergistic probability prediction method according to claim 9, characterized in that, The process of backpropagating the prediction bias to the auxiliary model through the cross-scale causal path in the time-factor feedback graph specifically involves: Key feature variables are identified based on prediction bias, and a set of cross-scale causal paths associated with the key feature variables is retrieved. Based on the causal path set, the sensitivity adjustment factor of each feature variable in the auxiliary model is calculated; The input features of each auxiliary model are subjected to perturbation simulation, and the potential residual contribution of each variable is calculated by weighting with sensitivity adjustment factor. The residual compensation tensor is constructed based on the potential residual contribution, perturbation amplitude, and causal strength.

Images