Multi-time scale charging load and energy storage cooperation probability prediction method

Abstract

The invention discloses a multi-time-scale charging load and energy storage cooperation probability prediction method, and relates to the technical field of power systems and smart grids, and the method comprises the steps: obtaining historical multi-source data of a target system, and constructing factor sub-graphs of different time scales; coupling the cause sub-graphs into a time cause feedback atlas through a structure mapping function, wherein the atlas comprises explicit time scale control nodes; executing cross-time-scale causal effect calculation based on the time factor feedback atlas to generate a causal effect matrix; using the causal effect matrix to construct a joint probability prediction model, and outputting a first prediction result; and when it is monitored that the deviation of the first prediction result exceeds a dynamic threshold value, triggering a time scale switching mechanism to generate a second prediction result. According to the method, multi-scale causal modeling is realized by constructing the time cause feedback atlas containing the time scale control nodes, dynamic scale switching and a residual compensation mechanism are combined, the prediction precision and the scheduling adaptability are remarkably improved, and multi-time-scale collaborative optimization management is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems and smart grids, and more specifically, to a method for predicting the probability of coordination between charging loads and energy storage at multiple time scales. Background Art

[0002] With the rapid growth of distributed power sources and charging loads in power systems, grid loads are exhibiting high volatility and multi-timescale characteristics. Charging loads, including electric vehicles and energy storage facilities, experience significant power demand fluctuations at minute, hour, and even daily levels. Energy storage systems regulate grid balance and improve system stability by storing and releasing energy at different timescales. Therefore, multi-timescale coordinated prediction of charging loads and energy storage has become an important research direction for load management and dispatch optimization in smart grids.

[0003] Existing technologies for multi-timescale load and energy storage forecasting primarily rely 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 the use of neural networks or random forest models to forecast energy storage charging and discharging. However, these methods often struggle to simultaneously capture the causal relationships and interactions across different timescales, resulting in limited forecast accuracy and a lack of dynamic adjustment mechanisms to address rapid changes in load and energy storage behavior. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a multi-time-scale charging load and energy storage collaborative probability prediction method. By constructing a time-induced feedback graph containing time-scale control nodes, structural mapping and coupling of causal subgraphs at different time scales are achieved, thereby improving the accuracy of load forecasting and the adaptability of scheduling, and realizing deep integration and refined collaborative management of multi-time-scale information.

[0005] To achieve the above object, the present invention provides the following technical solutions: A multi-time-scale charging load and energy storage collaborative probability prediction method includes the following steps: obtaining historical multi-source data of the target system and constructing causal subgraphs of different time scales; coupling the causal subgraphs into a time-induced feedback graph through a structural mapping function, wherein the graph includes explicit time-scale control nodes; performing cross-time-scale causal effect calculations based on the time-induced feedback graph to generate a causal effect matrix; using the causal effect matrix to construct a joint probability prediction model and output a first prediction result; and triggering a time-scale switching mechanism to generate a second prediction result when it is detected that the deviation of the first prediction result exceeds a dynamic threshold.

[0006] In a preferred embodiment, the method of acquiring historical multi-source data of the target system and constructing causal subgraphs of different time scales is as follows: acquiring historical multi-source data, performing time alignment through a dynamic time warping algorithm, and generating a synchronized time series data set; decomposing the data set using a hybrid algorithm of wavelet packet decomposition and empirical mode decomposition to obtain sub-data sets including short-term, medium-term, and long-term time scales; constructing an initial causal subgraph based on the sub-data sets through a combination of conditional independence test and correlation analysis; and performing structural optimization on the initial causal subgraph based on statistical significance, eliminating causal edges whose significance is lower than a preset threshold, and generating a final causal subgraph.

[0007] In a preferred embodiment, the causal subgraph is coupled into a temporal causal feedback graph through a structural mapping function, specifically: extracting node and edge information from short-term, medium-term, and long-term causal subgraphs to construct a causal variable space matrix of unified dimension; calculating the structural similarity and causal consistency of semantically identical nodes at multiple time scales based on the matrix to generate a node similarity matrix; establishing a node mapping index across time scales based on the node similarity matrix; based on the node mapping index, connecting the mapping nodes through cross-scale causal edges according to predefined cross-scale coupling rules to form a coupled causal graph; introducing a time scale control node in the coupled causal graph, and using the time scale control node as an intermediary to construct a cross-scale feedback path to generate a temporal causal feedback graph.

[0008] In a preferred embodiment, the method for screening time scale control nodes is as follows: based on the semantic and functional similarity of key feature variables in a multi-time scale causal subgraph, candidate time scale control nodes are screened; the comprehensive control index of the candidate nodes is calculated through a weighted evaluation model, and 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.

[0009] In a preferred embodiment, the causal effect calculation across time scales is performed based on the time-based feedback map to generate a causal effect matrix, specifically: key feature nodes and their causal paths under multiple time scales are extracted from the time-based feedback map to generate a causal path set; the initial causal effect of each path on the target variable is calculated through a causal inference algorithm to form an initial causal effect matrix; according to the directional relationship of the time-scale control node, the path causal weight is corrected and the confidence interval is adjusted to generate a final causal effect matrix; based on the final causal effect matrix, the time-based feedback map is incrementally corrected through a feedback mechanism.

[0010] In a preferred embodiment, the time-dependent feedback map is incrementally corrected through a feedback mechanism based on the final causal effect matrix, specifically: based on the final causal effect matrix, the causal contribution change rate of the key characteristic variables is calculated; when the change rate exceeds a preset change rate threshold, it is marked as a structural drift variable; the local causal structure incremental correction mechanism is triggered to adjust the topological relationship of the nodes and edges related to the drift variable; the minimum description length criterion is applied to select the optimal solution from the candidate adjustment solutions, and the corrected time-dependent feedback map is output.

[0011] In a preferred embodiment, the causal effect matrix is ​​used to construct a joint probability prediction model and output a 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-matrix; the corresponding probability prediction sub-model is constructed based on the direct causal effect sub-matrix of each time scale; the cross-scale indirect causal effect sub-matrix is ​​used to establish a feature projection channel to achieve short-term-medium-term-long-term step-by-step feature mapping; the long-term prediction result is back-propagated to the medium-term and short-term sub-models through feedback connection for correction; and the Bayesian model averaging method is used to fuse the outputs of the sub-models of each scale to generate the first prediction result.

[0012] In a preferred embodiment, when it is monitored that the deviation of the first prediction result exceeds a dynamic threshold, the time scale switching mechanism is triggered, specifically: the first prediction result is compared with the real-time power grid load monitoring data to calculate the prediction deviation of each time scale; when any scale deviation exceeds a preset threshold or the multi-scale prediction results conflict, the scale switching is triggered; a three-dimensional evaluation vector including the deviation change rate trend, the causal stability coefficient and the scheduling priority weight is constructed; the score of each scale is calculated by a preset evaluation function, and the scale with the highest score is selected as the optimal prediction scale; and the time scale switching instruction is output to the prediction system.

[0013] In a preferred embodiment, the generation of the second prediction result is specifically as follows: based on the time scale switching instruction, the model corresponding to the optimal time scale is selected as the main prediction model, and the remaining scale models are set as auxiliary models; the fusion weight coefficient of the main and auxiliary models is calculated based on the scheduling priority weight in the three-dimensional evaluation vector; the prediction deviation is back-propagated to the auxiliary model through the cross-scale causal path in the time-induced feedback graph; the residual compensation tensor is injected into the representation layer of the auxiliary model, the residual gradient is calculated and the weight is slightly updated; the main model output and the auxiliary model output are weighted and fused according to the weight coefficient to generate the second prediction result.

[0014] In a preferred embodiment, the prediction deviation is back-propagated to the auxiliary model through the cross-scale causal path in the time-dependent feedback map, specifically: identifying key feature variables based on the prediction deviation, and retrieving the cross-scale causal path set 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 in combination with the weighted sensitivity adjustment factor; and constructing a residual compensation tensor based on the potential residual contribution, perturbation amplitude and causal strength.

[0015] The technical effects and advantages of the multi-time-scale charging load and energy storage coordinated probability prediction method of the present invention are as follows: 1. By constructing multi-time-scale causal subgraphs and coupling them into time-dependent feedback graphs, the present invention 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, thereby achieving step-by-step mapping and correction of prediction results at each time scale. This significantly improves the accuracy of coordinated prediction of load and energy storage, and reduces prediction deviations caused by data noise or local fluctuations.

[0016] 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 and implements dynamic weight allocation and scale adaptive switching based on causal strength, providing a refined error correction path for multi-time scale 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. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the flow chart of the multi-time-scale charging load and energy storage coordination probability prediction method of the present invention; Figure 2 This is a schematic diagram of the second prediction result flow of the present invention. DETAILED DESCRIPTION

[0018] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] Example 1, Figure 1 The present invention provides a multi-time-scale charging load and energy storage coordination probability prediction method, including: S1, obtain historical multi-source data of the target system and construct causal subgraphs at different time scales; In this implementation case, historical multi-source data of the target system is obtained to construct causal subgraphs at different time scales, as follows: Historical multi-source data is collected through user-side apps, power companies, and meteorological service platforms. This includes, but is not limited to, charging load data from electric vehicle charging stations, charge and discharge status data from energy storage systems, grid operating parameters, user charging behavior records, and weather and holiday data. These data sources have different sampling frequencies and recording periods. To ensure temporal consistency in subsequent causal analysis, the data must first be time-aligned.

[0020] A dynamic time warping algorithm is used to align heterogeneous data from multiple sources to minimize nonlinear shifts along the time axis. Based on the global shortest distance criterion, this algorithm identifies the optimal alignment path between different time series and interpolates data at different moments onto a unified reference time axis, resulting in a structurally consistent and highly synchronized time series dataset that serves as the input for causal modeling.

[0021] After the time series data are synchronized, the dataset is scaled using a hybrid algorithm combining wavelet packet decomposition (WPD) and empirical mode decomposition (EMD). WPD is used to capture high-frequency transient features in the signal, while EMD is used to adaptively extract nonlinear trend components. This hybrid decomposition strategy effectively extracts sub-datasets at three time scales: short-term (daily), medium-term (weekly), and long-term (monthly). For each sub-dataset at each time scale, a causal structure learning method based on conditional independence tests and statistical correlation analysis is used to construct the corresponding causal subgraph. Specifically, an improved PC algorithm is used to perform conditional independence tests to identify independent relationships between variables under given conditions. Statistical correlation metrics (such as partial correlation coefficients and mutual information) are then combined to evaluate the significance and strength of edges, generating an initial causal subgraph for each time scale.

[0022] To ensure the statistical validity and engineering usability of the initial causal subgraph, the initial causal graph undergoes further structural verification. A significance screening mechanism is introduced to remove weakly correlated edges with weights below a set confidence threshold, preventing noise paths from interfering with the inference results. Finally, the final causal subgraph at each time scale, after significance screening, is output as the basic input for subsequent graph coupling and causal reasoning.

[0023] S2, coupling the causal subgraph into a temporal feedback graph through a structural mapping function, wherein the graph includes explicit time scale control nodes; In this implementation case, the causal subgraph is coupled into a time-cause feedback graph through a structural mapping function, specifically: First, we extract node and edge information from the short-term, medium-term, and long-term causal subgraphs. Each causal subgraph is defined as a directed graph: ,

[0024] in, is the set of causal variable nodes at the corresponding time scale, is the set of causal edges between nodes.

[0025] In order to achieve a unified representation of causal variables at different time scales, a causal variable representation space is designed to map variables at all time scales into a common feature space. This space is vectorized by encoding the semantic features, statistical features, and physical properties of nodes to form a feature vector for each node. ,in As the feature dimension, the feature vectors of all nodes at the same time scale are arranged in rows to form a causal variable space matrix :

[0026] in, Indicates the total number of nodes.

[0027] Based on the spatial matrix of causal variables at each time scale, the structural similarity and causal consistency between nodes with the same semantics or similar functions are calculated for subsequent node mapping.

[0028] Structural similarity is calculated using cosine similarity: , ,

[0029] in, and denote the short-term and medium-term time scales, respectively, and 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 the causal paths between nodes, which is defined as the similarity of the adjacency matrix of the nodes in their respective causal subgraphs, using the Jaccard similarity coefficient or path consistency index:

[0030] in, Represents the node adjacency set, integrating structural similarity and causal consistency to form a node similarity matrix:

[0031] in, is a hyperparameter for adjusting the weights of the two.

[0032] According to the node similarity matrix , using the maximum matching algorithm or threshold screening strategy to determine the one-to-one mapping relationship of nodes across time scales and generate a node mapping index set:

[0033] in, is the similarity threshold, which is used to eliminate low-similarity mappings to ensure mapping quality.

[0034] Node-based mapping index According to the predefined cross-scale coupling rules, the corresponding nodes on different time scales are connected through cross-scale causal edges to form a coupled causal graph ,in, , , is the set of cross-scale causal edges.

[0035] The screening method of the time scale control node is: 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 in the coupled causal graph as the intermediary hub of the feedback relationship between scales.

[0036] 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 foundation for selecting time-scale control nodes. All data must first be formatted uniformly, and feature variable nodes must be standardized to facilitate cross-scale comparisons.

[0037] The semantic similarity between node pairs is calculated and the functional type consistency is judged. Label matching is performed on the variable types (such as load type, behavior type, and environment type) to which the nodes belong. Nodes with inconsistent types are eliminated, and a set of nodes whose semantics and functions meet the threshold requirements are selected as the candidate time scale control node set.

[0038] Control node set for candidate time scale Each node in Calculate the following metrics: Causal Strength , node frequency characteristics and prediction sensitivity .

[0039] Causal Strength It represents the degree of causal influence between the node and the target predictor variable, and its calculation formula is;

[0040] in, is the set of time scales, Indicated on scale Next node Points to the target variable The edge weight of .

[0041] Node frequency characteristics It represents the frequency of the node appearing in the causal graph at each time scale, and its calculation formula is:

[0042] in, A certain time scale The causal subgraph below.

[0043] Prediction sensitivity It indicates the degree of influence of a small disturbance of the node input value on the prediction result, which can be calculated by partial derivative sensitivity analysis:

[0044] in, is the prediction model output, is the input node.

[0045] Causal intensity , node frequency characteristics and prediction sensitivity After the three indicators are normalized, the comprehensive score of each candidate node is calculated:

[0046] in, 、 、 are the normalized results of the corresponding indicators, 、 、 Weight parameters, satisfying , can be adjusted according to the prediction task (for example, when the priority is on prediction stability, A comprehensive score threshold is preset, and nodes with a comprehensive score greater than the comprehensive score threshold are used as the final time scale control node set. .

[0047] Cross-scale feedback edges are constructed with time-scale control nodes as intermediaries, connecting time-scale control nodes with causal nodes at each scale to form feedback paths, such as "short-term charging load → control node (total daily charging volume) → medium-term temperature" and "long-term charging trend → control node (monthly average temperature) → short-term temperature", etc., to achieve two-way flow and dynamic adjustment of causal information between different scales, improve the accuracy and stability of the prediction model, and finally output a time-cause feedback map containing time-scale control nodes and cross-scale feedback paths. ,in , , Cross-scale feedback edge aggregation.

[0048] S3, performs causal effect calculation across time scales based on the time-induced feedback graph to generate a causal effect matrix; In this implementation case, a causal effect calculation across time scales is performed based on the temporal feedback graph to generate a causal effect matrix, specifically: From the time-dependent feedback graph Key feature nodes and their causal paths at different time scales are extracted. Key feature nodes are nodes that significantly influence the target variable (such as charging load and energy storage status). They can be selected based on metrics such as node degree (number of connected edges) and causal strength. For example, "real-time charging power" and "instantaneous temperature" in the short term, "daily charging total" and "daily average temperature" in the medium term, "monthly charging total" and "monthly average temperature" in the long term, and time scale control nodes such as "daily average temperature" and "weekly charging total" are all selected as key feature nodes.

[0049] A causal path is a directed path from a key feature node to a target variable. Each path consists of a series of connected causal edges. Traverse the entire time-dependent feedback graph, collect all paths from key feature nodes to target variables, and output a set of causal paths. ,in, Indicates the causal path.

[0050] According to the causal path set , calculate the causal effect of the characteristic variables on the target variable on each path through causal inference algorithms (such as Do-calculus). For each causal path ,in, is the key feature variable, is the target variable, and its causal effect is the product of the weights of each edge on the path, and the formula is:

[0051] in, For the path The weight of the edge, For the same feature variable to the target variable The total causal effect of multiple paths is the sum of the causal effects of each path.

[0052] According to the above calculations, the initial causal effect matrix is ​​output ,in, is the number of key feature variables, is the number of target variables, the matrix elements Indicates the The key characteristic variables are The initial causal effect size of the target variable.

[0053] The directional relationship of the nodes is controlled according to the time scale, the path causal weight is corrected and the confidence interval is adjusted to generate the final causal effect matrix, which is specifically: The time scale control node has a regulatory effect on the node it points to. , if the path contains nodes controlled by time scale Pointing to the node , then the causal weight of the path needs to be multiplied by the control coefficient , the modified path causal effect formula is:

[0054] At the same time, the confidence interval of the causal effect is adjusted based on the influence of the control node. Let the initial confidence interval be:

[0055] in, is the standard deviation of the causal effect.

[0056] The revised confidence interval is:

[0057] in, .

[0058] Through the above corrections, the final causal effect matrix is ​​obtained .

[0059] The incremental correction of the time-induced feedback map is performed through the feedback mechanism based on the final causal effect matrix, specifically: Based on the final causal effect matrix , calculate the causal contribution change rate of each key characteristic variable at the current time. The causal contribution change rate refers to the relative change between the current causal effect and the causal effect at the previous moment, and the formula is:

[0060] in, For the previous moment The key characteristic variables are The ultimate causal effect of the target variable, For the The rate of change of the causal contribution of the key characteristic variables.

[0061] Set preset rate of change threshold , if the causal contribution change rate of any key characteristic variable , the variable is marked as a structural drift variable. Based on the structural drift variable, the local causal structure incremental correction mechanism is triggered to make topological adjustments to the relevant nodes and edges in the time-dependent feedback graph. The relevant nodes include the structural drift variable itself, the nodes directly connected to it, and the nodes indirectly connected through the causal path. Adjustment methods include adjusting edge weights, adding or deleting edges. For edges related to the structural drift variable, the weight is adjusted according to the change in the causal effect. The formula is:

[0062] Among them, the adjustment coefficient , which controls the extent of weight adjustment. If the causal effect between the structural drift variable and a node is significantly enhanced (exceeding the threshold for adding edges, such as 1.5 times the original effect), an edge is added. If the causal effect is significantly weakened (below the threshold for deleting edges, such as 0.5 times the original effect), the edge is deleted.

[0063] The optimal adjustment scheme is selected by the minimum description length criterion (MDL). The minimum description length criterion comprehensively considers the complexity of the model and the degree of fit to the data. The formula is:

[0064] in, is the structural encoding length of the adjusted time-dependent feedback graph (such as the encoding length of the number of nodes, the number of edges, etc.), The length of the fitting error encoding of the data based on the adjusted spectrum (the smaller the fitting error, the shorter the encoding length). Select the adjustment scheme with the smallest MDL value to generate the corrected time-dependent feedback spectrum .

[0065] S4, constructing a joint probability prediction model using the causal effect matrix and outputting the first prediction result; In this implementation case, a joint probability 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. The direct causal effect sub-matrix reflects the causal effect of the key characteristic variables on the target variable within the same time scale. The elements corresponding to the key characteristic variables and the target variables in the final causal effect matrix belonging to the same time scale are extracted to form the direct causal effect sub-matrix, including the short-term direct causal effect sub-matrix. , medium-term direct causal effect sub-matrix and the long-term direct causal effect submatrix .

[0066] The cross-scale indirect causal effect submatrix reflects the causal effect of key characteristic variables on the target variable at different time scales, including the short-term to medium-term cross-scale indirect causal effect submatrix and the medium- to long-term cross-scale indirect causal effect submatrix .

[0067] LSTM (Long Short-Term Memory Network) is used to build a short-term probability prediction sub-model, which takes short-term key feature variables as input and combines the short-term direct causal effect sub-matrix , the LSTM network is used to capture the short-term dependencies of the time series and output the probability distribution of the short-term target variable. The model output layer uses the softmax activation function to obtain the probability distribution ,in, is the short-term input feature, is the short-term target variable.

[0068] GRU (Gated Recurrent Unit) is used to build a medium-term probability prediction sub-model. The medium-term key feature variables and the medium-term direct causal effect sub-matrix are used. , GRU network can effectively process medium-length time series and output the probability distribution of medium-term target variables ,in, is the mid-term input feature, is the medium-term target variable.

[0069] A long-term probability forecast sub-model is constructed using ARIMA (Autoregressive Integrated Moving Average model). Combined with the long-term key characteristic variables and the long-term direct causal effect sub-matrix The ARIMA model outputs the probability distribution of the long-term target variable through operations such as stabilization, autoregression and moving average of the time series. ,in, is the long-term input feature, is the long-term target variable.

[0070] The cross-scale indirect causal effect submatrix is ​​used to establish a feature projection channel to achieve short-term-medium-term-long-term step-by-step feature mapping. , a feature projection channel is established through the fully connected layer to map the output features of the short-term sub-model to the medium-term scale. The mapping formula is:

[0071] in, is the output feature vector of the short-term sub-model, is the weight matrix (whose elements are associated), is the bias term, is the ReLU activation function, is the feature vector mapped to the medium-term scale.

[0072] Similarly, based on the medium- to long-term cross-scale indirect causal effect submatrix , establish a feature projection channel to map the output features of the medium-term sub-model to the long-term scale. The mapping formula is:

[0073] in, is the output feature vector of the mid-term sub-model, is the weight matrix, is the bias term, is the feature vector mapped to the long-term scale. The long-term prediction results are back-propagated to the medium-term and short-term sub-models through feedback connections to perform cross-scale prediction correction. The prediction results output by the long-term sub-model are The prediction results of the mid-term sub-model are corrected through the feedback channel. The correction formula is:

[0074] in, is the original prediction result of the mid-term sub-model, is the correction coefficient (determined based on the cross-scale causal effect), It is the prediction result of the medium-term sub-model based on the long-term characteristics.

[0075] Similarly, the adjusted mid-term forecast results It is then transmitted to the short-term sub-model through the feedback channel to correct the short-term prediction results.

[0076] The Bayesian model averaging method is used to fuse the outputs of the sub-models at each scale to obtain the first prediction result. The Bayesian model averaging method assigns different weights according to the prediction performance of each sub-model. The weight calculation formula is:

[0077] in, For the The negative log-likelihood value of each sub-model (reflects the goodness of fit of the model, the smaller the value, the better the fit), For the The weight of each sub-model, is the sum of the negative log-likelihood indexes of each sub-model, Corresponding to the short-term, medium-term and long-term sub-models respectively, For the Specifically, if the negative log-likelihood value of a sub-model is Smaller (i.e., better model fitting effect), The larger the value, the higher the proportion in the total, and the corresponding weight of the sub-model In this way, the weights can objectively reflect the prediction performance of each sub-model, making the fusion result more dependent on the model with better performance.

[0078] The probability distribution of the first prediction result after fusion is:

[0079] in, 、 、 are the probability distributions of the corrected short-term, medium-term, and long-term sub-models, respectively. is the target variable after fusion.

[0080] Through the above fusion, the first prediction result is obtained .

[0081] S5, when it is monitored that the deviation of the first prediction result exceeds a dynamic threshold, triggering a time scale switching mechanism to generate a second prediction result; In this implementation case, the time scale switching mechanism is triggered as follows: The first prediction result is compared with the real-time monitored grid load data to calculate the prediction deviation. The prediction deviation is measured using the root mean square error (RMSE) and the formula is:

[0082] in, The first prediction result The predicted value at the moment, The first one in the real-time monitoring of power grid load data The actual value at the moment, is the number of data samples. For each time scale, the forecast deviation is calculated for the short-term, medium-term, and long-term. The forecast deviation for each time scale is compared with the preset deviation threshold. If the forecast deviation at any time scale exceeds its corresponding preset deviation threshold, or if there is a conflict between the forecast results at multiple time scales (for example, a short-term forecast load increase, while the medium-term and long-term forecast load decreases without reasonable basis), a time scale switch is triggered.

[0083] Construct a three-dimensional evaluation vector that includes the deviation change rate trend, causal stability coefficient, and scheduling priority weight ,Deviation change rate trend Reflects the changing direction and speed of the forecast deviation. The calculation formula is:

[0084] in, is the prediction deviation at the current moment, is the prediction deviation of the previous moment. Causal stability coefficient The stability of the causal relationship in the feedback graph is measured based on the causal effect matrix. The formula is:

[0085] in, is the standard deviation of the final causal effect matrix, is the mean of the final causal effect matrix, The closer it is to 1, the more stable the causal relationship is. Scheduling priority weight It is set according to the grid dispatching demand to reflect the importance of prediction results at different time scales in dispatching.

[0086] Based on the three-dimensional evaluation vector, the scores of each scale are calculated using a preset evaluation function. The evaluation function uses a weighted summation method, and the formula is:

[0087] in, For the Time scale ( ) score, 、 、 is the weight coefficient, , 、 、 Respectively The deviation change rate trend, causal stability coefficient, and scheduling priority weight corresponding to each time scale are calculated. The scale with 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 scores the highest, the switching instruction is to focus on the short-term scale for prediction.

[0088] The second prediction result is specifically obtained as follows: Based on the time scale switching instruction, the model corresponding to the optimal time scale is selected as the main prediction model, and the remaining scale models are set as auxiliary models. For example, if the switching instruction is a short-term scale, the short-term probability prediction sub-model is the main prediction model, and the medium-term and long-term probability prediction sub-models are auxiliary models. The fusion weight coefficient of the main and auxiliary models is calculated based on the scheduling priority weight in the three-dimensional evaluation vector. Let the scheduling priority weight corresponding to the main model be , the scheduling priority weight corresponding to the auxiliary model is 、 , then the calculation formula of the fusion weight coefficient is:

[0089]

[0090] and satisfy .

[0091] Identify key feature variables based on prediction deviations (such as "temperature", "traffic flow" and other variables that cause large deviations) and retrieve the cross-scale causal path set associated with the key feature variables , based on a cross-scale causal path set Calculate the sensitivity adjustment factor of each characteristic variable in the auxiliary model , The larger the value, the more sensitive the characteristic variable is to the prediction deviation. Perform perturbation simulation on the input features of each auxiliary model (such as increasing or decreasing the characteristic value by 5%), weight the simulation results with the sensitivity adjustment factor, and calculate the potential residual contribution of each characteristic variable in the current prediction scenario:

[0092] in, is the characteristic variable The amount by which the prediction results change after the disturbance.

[0093] Based on potential residual contribution , the corresponding disturbance amplitude (e.g. 5% eigenvalue) and causal strength , construct the residual compensation tensor , inject the residual compensation tensor into the intermediate representation layer of the auxiliary model, calculate the residual gradient between the output of the intermediate representation layer, and update the intermediate layer weight slightly according to the residual gradient (the learning rate is set to 0.01), and output the adjusted auxiliary prediction result .

[0094] The prediction results of the main model based on the fusion weight coefficient and auxiliary prediction results Perform weighted fusion to generate the second prediction result:

[0095] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0096] The above embodiments may be implemented in whole or in part through software, hardware, firmware or any other combination. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0097] Those skilled in the art will appreciate that the modules and algorithm steps of each example 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 performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel 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.

[0098] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

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

[0100] Finally: 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 in the scope of protection of the present invention.

Claims

1. A multi-time scale charging load and energy storage coordination probability prediction method, characterized by: The steps include: Obtain historical multi-source data of the target system and construct causal subgraphs at different time scales; The causal subgraphs are coupled into a temporal feedback graph through a structural mapping function, wherein the graph includes explicit time scale control nodes; Perform causal effect calculations across time scales based on the temporal feedback graph to generate a causal effect matrix; Use the causal effect matrix to build a joint probability prediction model and output the first prediction result; When it is monitored that the deviation of the first prediction result exceeds a dynamic threshold, a time scale switching mechanism is triggered to generate a second prediction result.

2. The multi-time-scale charging load and energy storage coordination probability prediction method according to claim 1 is characterized in that: The method of obtaining historical multi-source data of the target system and constructing causal subgraphs at different time scales is as follows: Acquire historical multi-source data, align them in time using a dynamic time warping algorithm, and generate synchronized time series datasets; The hybrid algorithm of wavelet packet decomposition and empirical mode decomposition is used to decompose the data set to obtain sub-datasets including short-term, medium-term and long-term time scales; Based on the sub-dataset, an initial causal subgraph is constructed by combining a conditional independence test and a correlation analysis method; The initial causal subgraph is structurally optimized based on statistical significance, and causal edges with significance lower than a preset threshold are removed to generate the final causal subgraph.

3. The multi-time-scale charging load and energy storage coordination probability prediction method according to claim 2 is characterized in that: The causal subgraph is coupled into a time-dependent feedback graph through the structural mapping function, specifically: Extract node and edge information from short-term, medium-term, and long-term causal subgraphs and construct a causal variable space matrix with unified dimensions; Based on the matrix, the structural similarity and causal consistency of semantically identical nodes at multiple time scales are calculated to generate a node similarity matrix; According to the node similarity matrix, a node mapping index across time scales is established; Based on the node mapping index, according to the predefined cross-scale coupling rules, the mapping nodes are connected through cross-scale causal edges to form a coupled causal graph; Time scale control nodes are introduced into the coupled causal graph, and cross-scale feedback paths are constructed with time scale control nodes as intermediaries to generate a time-cause feedback graph.

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

5. The multi-time-scale charging load and energy storage coordination probability prediction method according to claim 4 is characterized in that: The causal effect calculation across time scales is performed based on the time-induced feedback graph to generate a causal effect matrix, specifically: Extract key feature nodes and their causal paths at multiple time scales from the temporal feedback graph to generate a causal path set; Calculate the initial causal effect of each path on the target variable through the causal inference algorithm to form an initial causal effect matrix; According to the time scale, the directional relationship of the control nodes is modified, the path causal weight is adjusted, and the confidence interval is adjusted to generate the final causal effect matrix; Based on the final causal effect matrix, the time-cause feedback map is incrementally corrected through the feedback mechanism.

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

7. The multi-time-scale charging load and energy storage coordination probability prediction method according to claim 6 is characterized in that: The causal effect matrix is ​​used to construct a joint probability prediction model and output a first prediction result, specifically: Decompose the final causal effect matrix to separate the short-term, medium-term, and long-term direct causal effect sub-matrices and the cross-scale indirect causal effect sub-matrix; Based on the direct causal effect sub-matrix of each time scale, the corresponding probability prediction sub-model is constructed; A feature projection channel is established using the cross-scale indirect causal effect submatrix to achieve short-term-medium-term-long-term step-by-step feature mapping; The long-term prediction results are back-propagated to the medium-term and short-term sub-models through feedback connections for correction; The Bayesian model averaging method is used to fuse the outputs of sub-models at each scale to generate the first prediction result.

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

9. The multi-time-scale charging load and energy storage coordination probability prediction method according to claim 8 is characterized in that: The generating of the second prediction result is specifically as follows: Based on the time scale switching instruction, the model corresponding to the optimal time scale is selected as the main prediction model, and the remaining scale models are set as auxiliary models; Calculate the fusion weight coefficient of the main and auxiliary models based on the scheduling priority weight in the three-dimensional evaluation vector; Backpropagating prediction biases to auxiliary models via cross-scale causal paths in the temporal feedback graph; Inject the residual compensation tensor into the representation layer of the auxiliary model, calculate the residual gradient and update the weights slightly; The output of the main model and the output of the auxiliary model are weighted and fused according to the weight coefficient to generate the second prediction result.

10. The multi-time-scale charging load and energy storage coordination probability prediction method according to claim 9 is characterized in that: The prediction bias is back-propagated to the auxiliary model through the cross-scale causal path in the time-dependent feedback graph, specifically: Identify key feature variables based on prediction bias and retrieve a set of cross-scale causal paths associated with the key feature variables; Based on the set of causal paths, the sensitivity adjustment factor of each characteristic variable in the auxiliary model is calculated; Perform disturbance simulation on the input features of each auxiliary model, and calculate the potential residual contribution of each variable by weighting it with the sensitivity adjustment factor; The residual compensation tensor is constructed based on the potential residual contribution, disturbance amplitude and causal strength.

Images