A Power Load Dispatch Method and System Based on ARIMAX Model and Causal Forest

Abstract

This invention discloses a power load dispatching method and system based on the ARIMAX model and causal forest, comprising the following steps: collecting dispatching modeling data of the target dispatching area and completing data preprocessing to generate a preprocessed dispatching sample sequence; training the ARIMAX baseline load prediction model to obtain a baseline load prediction sequence; performing residual incremental target construction to generate a causal training dataset; performing time-series block honest training on the causal training dataset to obtain an honest causal forest model; inputting the preprocessed dispatching sample sequence into the honest causal forest model to output a treatment effect estimation sequence and a treatment effect confidence interval; performing fusion operations to obtain a policy-conditional load prediction sequence; performing lower bound constraint dispatching of the effect interval, selecting dispatching action execution objects based on the lower bound, calculating the dispatching action allocation amount, and generating a dispatching instruction set. This invention improves the allocation reliability of power load dispatching.

Description

Technical Field

[0001] This invention relates to the field of intelligent power system dispatching technology, and in particular to a power load dispatching method and system based on the ARIMAX model and causal forest. Background Technology

[0002] With the increasing proportion of renewable energy grid connection, the deepening of electricity market transactions, and the widespread adoption of controllable loads and energy storage resources on the user side, the power system is placing higher demands on the accuracy, robustness, and real-time performance of load forecasting and dispatching decisions. Existing power load dispatching typically relies on historical load time series as a basis, combined with exogenous factors such as weather, holidays, and time-of-use pricing to forecast loads, and then formulates dispatching schemes such as demand response, load shedding, or energy storage charging and discharging based on this forecast.

[0003] However, traditional prediction-scheduling processes still have shortcomings when facing the feedback impact of scheduling actions on the load. Many solutions treat scheduling actions as exogenous inputs or static parameters, and the scheduling effect is often estimated through empirical coefficients or offline regression. It is difficult to characterize the heterogeneous impact of scheduling actions under different users, different scenarios, and different time periods. This leads to problems such as unstable effects, over- or under-response in the actual execution of scheduling action allocation. Even when using models such as random forests and gradient boosting trees to learn the "action-load" relationship, they are mostly correlation modeling, which makes it difficult to reliably estimate the true processing effect of scheduling actions when there are confounding factors. As a result, it is difficult to form a credible lower bound that can be used to constrain scheduling.

[0004] Most dispatching schemes lack a mechanism to explicitly introduce "effect uncertainty" into the command generation process. They usually only rely on point estimates or average effects to select targets and calculate allocation quantities. They lack conservative constraints and reachability checks based on the lower bound of the confidence interval. As a result, when load fluctuations intensify or exogenous variables change abruptly, dispatching commands cannot balance reliability and executability, affecting the safe operation of the power grid and the achievement of dispatching objectives.

[0005] Therefore, how to provide a power load scheduling method and system based on the ARIMAX model and causal forest is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a power load dispatching method and system based on the ARIMAX model and causal forest. This invention constructs a baseline load forecast based on ARIMAX and outputs the processing effect estimation sequence and its confidence interval of dispatching actions through an honest causal forest. The lower bound of the effect interval is introduced into the screening of dispatching objects and the calculation of allocation, thereby constraining and controlling the uncertainty of dispatching effect, reducing the probability of over- or under-dispatch, and improving the reliability, executability and operational safety of dispatching allocation.

[0007] The power load dispatching method based on the ARIMAX model and causal forest according to embodiments of the present invention includes the following steps:

[0008] Collect scheduling modeling data for the target scheduling area, complete data preprocessing, and generate a preprocessed scheduling sample sequence;

[0009] The ARIMAX baseline load prediction model is trained using preprocessed scheduling sample sequences to predict the target scheduling period, thus obtaining the baseline load prediction sequence for the target scheduling period.

[0010] Based on the preprocessed scheduling sample sequence and the baseline load prediction sequence, residual incremental target construction is performed, incremental load sample sequence is calculated, and causal training dataset is generated by merging them.

[0011] Honest training is performed on the causal training dataset by time-series block division, which is divided into a structure learning subset and an effect estimation subset. The structure learning subset is used to generate a tree structure, and the effect estimation subset is used for training to obtain an honest causal forest model.

[0012] Input the preprocessed scheduling sample sequence corresponding to the target scheduling period into the honest causal forest model, perform honest causal forest effect inference and interval generation, and output the treatment effect estimation sequence and treatment effect confidence interval;

[0013] A fusion operation is performed based on the treatment effect estimation sequence and the baseline load forecast sequence to obtain the policy-conditional load forecast sequence;

[0014] The lower bound constraint scheduling of the effect interval is applied to the load prediction sequence and the confidence interval of the treatment effect. The scheduling action execution objects are selected based on the lower bound, the scheduling action allocation is calculated, and a scheduling instruction set is generated.

[0015] Optionally, obtaining the preprocessed scheduling sample sequence specifically includes:

[0016] The scheduling modeling data of the power system in the target scheduling area is collected, and the scheduling modeling data is organized into a set of records arranged by sampling time according to a unified data recording format to obtain the original record sequence. Each record in the record set contains the load observation value, exogenous variables, scheduling actions and user or equipment profile information corresponding to the sampling time and the current sampling time.

[0017] Data preprocessing is performed on the original record sequence, including timestamp unification, time granularity alignment, missing value imputation, outlier handling, and time period feature construction, and a preprocessed scheduling sample sequence is generated.

[0018] Optionally, obtaining the baseline load forecast sequence for the target scheduling period specifically includes:

[0019] Load observations are extracted from the preprocessed scheduling sample sequence in the order of sampling time to form a load time series, and exogenous variables corresponding to each sampling time are extracted simultaneously to form an exogenous variable sequence, and an aligned data table of sampling time-load observation-exogenous variable is established.

[0020] Differential processing is performed on the load time series. The difference between the multi-order historical load observations is calculated according to the preset differential order to obtain the current differential value. The final differential values ​​corresponding to each sampling time are arranged in chronological order to obtain the differential load series. In the aligned data table, the differential load series replaces the original load time series as the modeling sequence, while retaining the end historical load observations before differential processing.

[0021] An ARIMAX model is built and trained on the differential load sequence. Candidate combinations of autoregressive and moving average orders are set, and corresponding model input structures are constructed based on exogenous variables in the preprocessed scheduling sample sequence. Parameter estimation is performed on each candidate order combination, the corresponding model fitting error is calculated, and the order combination with the smallest fitting error is selected as the final model structure. After determining the model structure, parameter estimation is performed on the autoregressive coefficients, moving average coefficients, and exogenous variable coefficients, and model training is completed. The exogenous variable sequence corresponding to the target scheduling period is input into the trained ARIMAX model, and the differential prediction sequence of the target scheduling period is generated recursively in time order. Combined with the retained historical load observations at the front and rear of the differential sequence, inverse differential reconstruction is performed on the differential prediction sequence time by time to obtain the baseline load prediction sequence of the target scheduling period.

[0022] Optionally, obtaining the causal training dataset specifically includes:

[0023] Extract data segments from the preprocessed scheduling sample sequence that cover the same target scheduling period as the baseline load prediction sequence, and align them with the baseline load prediction sequence time by time to form an aligned record sequence;

[0024] The residualized incremental target is constructed time-by-time for the aligned record sequence. The load observation value at each sampling time is subtracted from the baseline load prediction value at the same sampling time to obtain the incremental load sample value at the current sampling time. The incremental load sample sequence is obtained by arranging them in chronological order.

[0025] The incremental load sample sequence is merged with the exogenous variables, scheduling actions, and user or device profiles in the aligned record sequence according to the sampling time to generate a causal training dataset.

[0026] Optionally, the honest causal forest model is obtained by specifically including:

[0027] The causal training dataset is divided into time-series blocks according to the sampling time sequence. The block length is set to the preset number of sampling points. Starting from the initial sampling time, the dataset is divided into continuous time blocks according to the block length, and each time block is assigned a block number.

[0028] Honest partitioning is performed on each time block. A subset of time blocks are selected to form a structure learning subset, and the remaining time blocks that do not overlap with the structure learning subset are selected to form an effect estimation subset. The set of block numbers corresponding to the structure learning subset and the effect estimation subset is recorded.

[0029] By using the structure learning subset to generate tree structure, sample subsampling is performed on each tree. Candidate splitting features are selected from the covariate field within each node to be split, and candidate splitting thresholds are traversed. For each candidate splitting threshold, the node sample is divided into left and right child nodes. At the same time, the splitting score is calculated based on the difference in the processing effect of the left and right child nodes. The candidate splitting feature with the highest splitting score and the candidate splitting threshold are selected.

[0030] Repeat the node splitting process for each tree until the preset stopping condition is met, and obtain the tree structure set determined by the structure learning subset;

[0031] Leaf node effects are estimated using a subset of effect estimates. Samples in the subset of effect estimates are assigned to corresponding leaf nodes according to the tree structure path. Within each leaf node, the mean of the result field of the treated state sample and the mean of the result field of the untreated state sample are calculated and subtracted to obtain the estimated value of the treatment effect of the leaf node. The estimated values ​​of the treatment effect of the leaf nodes of each tree are encapsulated with the tree structure set to form an honest causal forest model.

[0032] Optionally, the output of the treatment effect estimation sequence and the treatment effect confidence interval specifically includes:

[0033] Read the preprocessed scheduling sample sequence corresponding to the target scheduling period, extract the covariate field and scheduling action field of each sampling time in the order of sampling time, construct the model input sample of the sampling time, and write the model input sample into the inference queue;

[0034] The model input samples are sequentially retrieved from the inference queue and input into each tree of the honest causal forest model. The splitting features and splitting thresholds of the current node are read in each tree. Layer-by-layer comparison and branch selection are completed until the leaf node is reached. The leaf node identifier on each tree at the sampling time is recorded, and the effective sample quantity identifier and leaf node stability identifier of the leaf node are read simultaneously.

[0035] For each tree, the leaf node output leaf node identifier is filtered for validity. Leaf node results with a valid sample size less than a preset minimum threshold or stability lower than a preset threshold are removed. The removed leaf node results are backtracked and replaced with their parent nodes. After the filtering and backtracking are completed, the pre-stored leaf node processing effect estimate value of the corresponding leaf node is read to form a cross-tree processing effect estimate value set.

[0036] Weighted robust aggregation is performed on the cross-tree treatment effect estimate set. For each treatment effect estimate in the set, the corresponding effective sample size identifier is read as the weight. The treatment effect estimate set is truncated according to the weight to remove extreme values ​​of a preset proportion. At the same time, the remaining treatment effect estimates are weighted and averaged to obtain the treatment effect estimate at the current sampling time. The results are then arranged in chronological order to generate the treatment effect estimate sequence.

[0037] Interval generation is performed on the set of cross-tree treatment effect estimates. The treatment effect estimates in the set are sorted by weight. The lower and upper quantile values ​​corresponding to the preset confidence level are extracted to obtain the original confidence interval. At the same time, consistency shrinkage is performed on the original confidence interval to remove trees that deviate from the treatment effect estimates by more than a preset threshold. The upper and lower bounds are then output and recalculated.

[0038] The confidence interval will be subjected to time consistency correction to eliminate abrupt changes between adjacent time intervals, and the confidence interval of the processing effect will be output.

[0039] Optionally, obtaining the strategy conditional load prediction sequence specifically includes:

[0040] Read the baseline load prediction sequence, treatment effect estimation sequence and treatment effect confidence interval of the target scheduling period, check the time index at each sampling time and complete the filling and elimination, and calculate the confidence weight corresponding to the width of the treatment effect confidence interval for each sampling time.

[0041] The target scheduling cycle is traversed in the order of sampling time. The estimated value of the treatment effect is scaled according to the confidence weight to obtain the corrected treatment effect value. At the same time, the corrected treatment effect value is restricted to the upper and lower bounds of the confidence interval of the treatment effect at the current sampling time to obtain the amplitude-limited treatment effect value, which is added to the baseline load forecast value to obtain the policy condition load forecast value.

[0042] The policy condition load prediction values ​​obtained at each sampling time are written into the result sequence in chronological order to form the policy condition load prediction sequence.

[0043] Optionally, the generation of the scheduling instruction set specifically includes:

[0044] Read the policy condition load prediction sequence and the confidence interval of the treatment effect. Extract the lower bound of the confidence interval at each sampling time to form a lower bound sequence. Perform median filtering on the lower bound sequence to suppress abrupt change points. Perform back-down replacement on the lower bound value that shows a reverse jump in a continuous time period to obtain a conservative lower bound sequence.

[0045] For each sampling time, the current lower bound value of the conservative lower bound sequence is read, and objects that have not reached the conservative threshold are removed to obtain a set of candidate objects. The conservative contribution density of each candidate object is calculated and sorted in descending order of conservative contribution density to form a sorted object list. The sorted object list is divided into the main execution layer and the supplementary layer.

[0046] For each sampling time, starting from the main execution layer, read the lower limit and upper limit of executable allocation of each object, generate the initial allocation of objects based on the remaining scheduling target quantity, perform a limit on the initial allocation of objects to obtain the object allocation quantity, and write the object allocation quantity into the allocation result set.

[0047] For each sampling time, the total lower bound guaranteed contribution is obtained by summarizing the allocation result set and comparing it with the scheduling target amount. When the total lower bound guaranteed contribution is insufficient, objects are added from the supplementary layer and the allocation amount calculation is repeated until the target amount is met or there are no available objects. The scheduling target amount is then reduced by a preset conservative ratio and the allocation result set is recalculated.

[0048] The allocation results of each sampling time within the target scheduling period are summarized in chronological order to generate a scheduling instruction set.

[0049] According to an embodiment of the present invention, a power load dispatching system based on an ARIMAX model and causal forest includes:

[0050] The data acquisition and preprocessing module is used to acquire scheduling modeling data of the target scheduling area, perform data processing, and generate preprocessed scheduling sample sequences.

[0051] The ARIMAX baseline prediction module is used to receive preprocessed scheduling sample sequences and train the ARIMAX baseline load prediction model, predict the target scheduling period, and output the baseline load prediction sequence.

[0052] The residual incremental construction module is used to receive the preprocessed scheduling sample sequence and the baseline load prediction sequence, subtract them at each time step to generate the incremental load sample sequence, and merge them to generate the causal training dataset.

[0053] The Honest Causal Forest Training Module is used to receive the causal training dataset and divide it into a structure learning subset and an effect estimation subset according to time blocks, generate a tree structure and train the Honest Causal Forest Model.

[0054] The effect inference and interval generation module is used to receive the honest causal forest model, locate the leaf nodes and read the leaf node effects, and aggregate the output of the processed effect estimation sequence and the processed effect confidence interval.

[0055] The fusion prediction and scheduling instruction generation module is used to perform effect fusion to obtain the policy condition load prediction sequence, filter the execution objects, calculate the allocation amount, and generate a scheduling instruction set.

[0056] The beneficial effects of this invention are:

[0057] Based on the preprocessing of the scheduling modeling data of the target scheduling area, this invention trains the ARIMAX baseline load prediction model to obtain the baseline load prediction sequence. By constructing residual incremental targets, the incremental part of the load change that can be attributed to scheduling actions is separated from the baseline, so that subsequent modeling can focus on the real impact of scheduling actions on the load, thereby improving the consistency and usability between load prediction and scheduling decision-making.

[0058] This invention performs temporal block honest training on the causal training dataset to obtain an honest causal forest model, and outputs the estimated sequence of treatment effects and the confidence interval of treatment effects within the target scheduling period. This enables the estimation of scheduling action effects to remain stable in the temporal scenario and explicitly gives the uncertainty range, thereby reducing the risk of effect bias caused by traditional correlation modeling.

[0059] This invention introduces the lower bound of the confidence interval of the processing effect into the scheduling process. By using the lower bound constraint, the scheduling action execution object is selected and the allocation amount is calculated, forming a scheduling instruction set for implementation. This enables the scheduling decision to have a conservative and controllable safety boundary and higher allocation reliability, thereby reducing the target deviation caused by over-scheduling or under-scheduling, and improving the stability of scheduling execution and the safety of power grid operation. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of the power load scheduling method based on the ARIMAX model and causal forest proposed in this invention;

[0062] Figure 2 This is a schematic diagram of the power load dispatching system based on the ARIMAX model and causal forest proposed in this invention.

[0063] Figure 3 This is a schematic diagram of the time-series block honest causal effect modeling and interval inference of the power load scheduling method based on the ARIMAX model and causal forest proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figures 1-3 The power load dispatching method based on the ARIMAX model and causal forest includes the following steps:

[0066] Collect scheduling modeling data for the target scheduling area, complete data preprocessing, and generate a preprocessed scheduling sample sequence;

[0067] The ARIMAX baseline load prediction model is trained using preprocessed scheduling sample sequences to predict the target scheduling period, thus obtaining the baseline load prediction sequence for the target scheduling period.

[0068] Based on the preprocessed scheduling sample sequence and the baseline load prediction sequence, residual incremental target construction is performed, incremental load sample sequence is calculated, and causal training dataset is generated by merging them.

[0069] Honest training is performed on the causal training dataset by time-series block division, which is divided into a structure learning subset and an effect estimation subset. The structure learning subset is used to generate a tree structure, and the effect estimation subset is used for training to obtain an honest causal forest model.

[0070] Input the preprocessed scheduling sample sequence corresponding to the target scheduling period into the honest causal forest model, perform honest causal forest effect inference and interval generation, and output the treatment effect estimation sequence and treatment effect confidence interval;

[0071] A fusion operation is performed based on the treatment effect estimation sequence and the baseline load forecast sequence to obtain the policy-conditional load forecast sequence;

[0072] The lower bound constraint scheduling of the effect interval is applied to the load prediction sequence and the confidence interval of the treatment effect. The scheduling action execution objects are selected based on the lower bound, the scheduling action allocation is calculated, and a scheduling instruction set is generated.

[0073] In this embodiment, obtaining the preprocessed scheduling sample sequence specifically includes:

[0074] The scheduling modeling data of the power system in the target scheduling area is collected and organized into a set of records arranged by sampling time according to a unified data recording format to obtain the original record sequence. Each record in the set contains the load observation value, exogenous variables, scheduling actions and user or equipment profile information corresponding to the sampling time and the current sampling time. The load observation value is used to characterize the actual electricity demand at each sampling time. The exogenous variables are used to characterize the external impact of temperature, electricity price, holidays, etc. on the load. The scheduling actions are used to characterize the control or response measures issued by the system at each time. The user or equipment profile information is used to characterize the electricity consumption characteristics of different users or equipment to support cluster modeling and differentiated scheduling.

[0075] Data preprocessing is performed on the original record sequence. The data preprocessing includes timestamp unification, time granularity alignment, missing value imputation, outlier handling, and time period feature construction. Timestamp unification includes converting the time representations of different data sources to the same time zone and the same time format and unifying them to the sampling time field. Time granularity alignment includes resampling the original record sequence to a unified sampling interval and aligning it to the same time grid. Missing value imputation includes performing imputation operations on missing load observations, exogenous variables, scheduling actions, or user or equipment profile information on the time grid to form a gap-free sequence. Outlier handling includes calculating the local central value and local dispersion of load observations within a preset sliding window and performing truncation replacement or neighborhood replacement on load observations that exceed a preset threshold. Time period feature construction includes generating time period indicator features based on the sampling time and appending them to the records of the corresponding time, and generating a preprocessed scheduling sample sequence.

[0076] In this embodiment, obtaining the baseline load forecast sequence for the target scheduling period specifically includes:

[0077] Load observations are extracted from the preprocessed scheduling sample sequence in the order of sampling time to form a load time series, and exogenous variables corresponding to each sampling time are extracted simultaneously to form an exogenous variable sequence, and an aligned data table of sampling time-load observation-exogenous variable is established.

[0078] Differential processing is performed on the load time series. The difference between the multi-order historical load observations is calculated according to the preset differential order to obtain the current differential value. The final differential values ​​corresponding to each sampling time are arranged in chronological order to obtain the differential load series. In the aligned data table, the differential load series replaces the original load time series as the modeling sequence, while retaining the end historical load observations before differential processing.

[0079] An ARIMAX model is built and trained on the differential load sequence. Candidate combinations of autoregressive and moving average orders are set, and corresponding model input structures are constructed based on exogenous variables in the preprocessed scheduling sample sequence. Parameter estimation is performed on each candidate order combination, the corresponding model fitting error is calculated, and the order combination with the smallest fitting error is selected as the final model structure. After determining the model structure, parameter estimation is performed on the autoregressive coefficients, moving average coefficients, and exogenous variable coefficients, and model training is completed. The exogenous variable sequence corresponding to the target scheduling period is input into the trained ARIMAX model, and the differential prediction sequence of the target scheduling period is generated recursively in time order. Combined with the retained historical load observations at the front and rear of the differential sequence, inverse differential reconstruction is performed on the differential prediction sequence time by time to obtain the baseline load prediction sequence of the target scheduling period.

[0080] This invention aligns the load with exogenous variables over time, trains an ARIMAX model after differential stabilization of the load sequence, selects the optimal model structure from candidate order combinations and completes parameter estimation, then recursively predicts the target scheduling period and performs inverse differential reconstruction to obtain the baseline load prediction sequence. This improves the baseline prediction's ability to characterize time dependence and external drivers, reduces the interference of non-stationary fluctuations on the prediction, and enhances the accuracy and stability of the baseline load for the target scheduling period, providing a reliable benchmark for incremental construction and scheduling effect evaluation.

[0081] In this embodiment, obtaining the causal training dataset specifically includes:

[0082] Extract data segments from the preprocessed scheduling sample sequence that cover the same target scheduling period as the baseline load prediction sequence, and align them with the baseline load prediction sequence time by time to form an aligned record sequence;

[0083] The residualized incremental target is constructed time-by-time for the aligned record sequence. The load observation value at each sampling time is subtracted from the baseline load prediction value at the same sampling time to obtain the incremental load sample value at the current sampling time. The incremental load sample sequence is obtained by arranging them in chronological order.

[0084] The incremental load sample sequence is merged with the exogenous variables, scheduling actions, and user or device profiles in the aligned record sequence according to the sampling time to generate a causal training dataset.

[0085] In this embodiment, the honest causal forest model is obtained specifically through the following steps:

[0086] The causal training dataset is divided into time-series blocks according to the sampling time sequence. The block length is set to the preset number of sampling points. Starting from the initial sampling time, the dataset is divided into continuous time blocks according to the block length, and each time block is assigned a block number.

[0087] Honest partitioning is performed on each time block. A subset of time blocks are selected to form a structure learning subset, and the remaining time blocks that do not overlap with the structure learning subset are selected to form an effect estimation subset. The set of block numbers corresponding to the structure learning subset and the effect estimation subset is recorded.

[0088] By using the structure learning subset to generate tree structure, sample subsampling is performed on each tree. Candidate splitting features are selected from the covariate field within each node to be split, and candidate splitting thresholds are traversed. For each candidate splitting threshold, the node sample is divided into left and right child nodes. At the same time, the splitting score is calculated based on the difference in the processing effect of the left and right child nodes. The candidate splitting feature with the highest splitting score and the candidate splitting threshold are selected.

[0089] Repeatedly perform node splitting on each tree until a preset stopping condition is met. The stopping condition includes any one of the following: reaching a preset maximum tree depth, the number of leaf node samples being less than a preset minimum number of leaf node samples, or the splitting score being lower than a preset minimum splitting score threshold. This yields a set of tree structures determined by the structure learning subset.

[0090] Leaf node effects are estimated using a subset of effect estimates. Samples in the subset of effect estimates are assigned to corresponding leaf nodes according to the tree structure path. Within each leaf node, the mean of the result field of the treated state sample and the mean of the result field of the untreated state sample are calculated and subtracted to obtain the estimated value of the treatment effect of the leaf node. The estimated values ​​of the treatment effect of the leaf nodes of each tree are encapsulated with the tree structure set to form an honest causal forest model.

[0091] This invention achieves independence between tree structure generation and leaf node effect estimation by temporally partitioning the causal training dataset and implementing honest partitioning of structure learning and effect estimation. In the structure learning stage, splitting selection is driven by the difference in processing effects to form a stable and generalizable set of tree structures. In the effect estimation stage, the processing effects of leaf nodes are calculated based on independent samples. This reduces the risk of adaptive bias and temporal leakage of homogeneous data, improves the robustness of processing effect estimation and the credibility of interval inference, and provides a reliable basis for subsequent lower bound constraint scheduling.

[0092] In this embodiment, the outputs of the treatment effect estimation sequence and the treatment effect confidence interval specifically include:

[0093] Read the preprocessed scheduling sample sequence corresponding to the target scheduling period, extract the covariate field and scheduling action field of each sampling time in the order of sampling time, construct the model input sample of the sampling time, and write the model input sample into the inference queue;

[0094] The model input samples are sequentially retrieved from the inference queue and input into each tree of the honest causal forest model. The splitting features and splitting thresholds of the current node are read in each tree. Layer-by-layer comparison and branch selection are completed until the leaf node is reached. The leaf node identifier on each tree at the sampling time is recorded, and the effective sample quantity identifier and leaf node stability identifier of the leaf node are read simultaneously.

[0095] For each tree, the leaf node output leaf node identifier is filtered for validity. Leaf node results with a valid sample size less than a preset minimum threshold or stability lower than a preset threshold are removed. The removed leaf node results are backtracked and replaced with their parent nodes. After the filtering and backtracking are completed, the pre-stored leaf node processing effect estimate value of the corresponding leaf node is read to form a cross-tree processing effect estimate value set.

[0096] Weighted robust aggregation is performed on the cross-tree treatment effect estimate set. For each treatment effect estimate in the set, the corresponding effective sample size identifier is read as the weight. The treatment effect estimate set is truncated according to the weight to remove extreme values ​​of a preset proportion. At the same time, the remaining treatment effect estimates are weighted and averaged to obtain the treatment effect estimate at the current sampling time. The results are then arranged in chronological order to generate the treatment effect estimate sequence.

[0097] Interval generation is performed on the set of cross-tree treatment effect estimates. The treatment effect estimates in the set are sorted by weight. The lower and upper quantile values ​​corresponding to the preset confidence level are extracted to obtain the original confidence interval. At the same time, consistency shrinkage is performed on the original confidence interval to remove trees that deviate from the treatment effect estimates by more than a preset threshold. The upper and lower bounds are then output and recalculated.

[0098] The confidence interval will be subjected to time consistency correction to eliminate abrupt changes between adjacent time intervals, and the confidence interval of the processing effect will be output.

[0099] In the inference stage, this invention suppresses effect drift caused by small sample leaf nodes by locating leaf nodes one by one and introducing effective sample size and stability screening of leaf nodes, as well as a parent node backoff mechanism. It also performs truncated weighted aggregation with effective sample size as weight to reduce interference from extreme tree outputs. At the same time, it extracts confidence intervals based on weighted quantiles and performs consistency shrinkage and temporal consistency correction to obtain a more robust treatment effect estimation sequence and a more reliable treatment effect confidence interval. This provides reliable input for lower bound constraint scheduling and improves the safety margin and executability of scheduling decisions.

[0100] In this embodiment, obtaining the strategy conditional load prediction sequence specifically includes:

[0101] Read the baseline load prediction sequence, treatment effect estimation sequence and treatment effect confidence interval of the target scheduling period, check the time index at each sampling time and complete the filling and elimination, and calculate the confidence weight corresponding to the width of the treatment effect confidence interval for each sampling time.

[0102] The target scheduling cycle is traversed in the order of sampling time. The estimated value of the treatment effect is scaled according to the confidence weight to obtain the corrected treatment effect value. At the same time, the corrected treatment effect value is restricted to the upper and lower bounds of the confidence interval of the treatment effect at the current sampling time to obtain the amplitude-limited treatment effect value, which is added to the baseline load forecast value to obtain the policy condition load forecast value.

[0103] The policy condition load prediction values ​​obtained at each sampling time are written into the result sequence in chronological order to form the policy condition load prediction sequence.

[0104] In this embodiment, the generation of the scheduling instruction set specifically includes:

[0105] Read the policy condition load prediction sequence and the confidence interval of the treatment effect. Extract the lower bound of the confidence interval at each sampling time to form a lower bound sequence. Perform median filtering on the lower bound sequence to suppress abrupt change points. Perform back-down replacement on the lower bound value that shows a reverse jump in a continuous time period to obtain a conservative lower bound sequence.

[0106] For each sampling time, the current lower bound value of the conservative lower bound sequence is read, and objects that have not reached the conservative threshold are removed to obtain a set of candidate objects. The conservative contribution density of each candidate object is calculated and sorted in descending order of conservative contribution density to form a sorted object list. The sorted object list is divided into the main execution layer and the supplementary layer.

[0107] For each sampling time, the executable allocation lower limit and executable allocation upper limit of each object are read one by one from the main execution layer. The initial allocation amount of the object is generated according to the remaining scheduling target amount. The initial allocation amount of the object is subjected to a limit to obtain the object allocation amount. The object allocation amount is written into the allocation result set. The lower bound guarantee contribution of the object is calculated based on the object allocation amount and the conservative lower bound value of the object to update the remaining scheduling target amount.

[0108] For each sampling time, the total lower bound guaranteed contribution is obtained by summarizing the allocation result set and comparing it with the scheduling target amount. When the total lower bound guaranteed contribution is insufficient, objects are added from the supplementary layer and the allocation amount calculation is repeated until the target amount is met or there are no available objects. The scheduling target amount is then reduced by a preset conservative ratio and the allocation result set is recalculated.

[0109] The allocation results of each sampling time within the target scheduling period are summarized in chronological order to generate a scheduling instruction set. The scheduling instruction set is used to implement corresponding scheduling actions on the selected execution objects according to the sampling time within the target scheduling period. It includes the execution object identifier, sampling time, scheduling action type and scheduling action allocation amount. The calculated scheduling action allocation results are converted into instructions that can be directly executed by the control side to complete scheduling objectives such as load reduction, peak shifting or energy storage charging and discharging.

[0110] This invention uses median filtering and backoff replacement to form a conservative lower bound sequence for the lower bound of the confidence interval of the processing effect. It then uses conservative thresholds and conservative contribution densities to screen, sort, and schedule execution objects in a hierarchical manner. In the allocation calculation, it combines the upper and lower bounds of executability to limit the range and updates the remaining target quantity in a closed loop with the contribution amount. At the same time, when the reachability is insufficient, it completes the repair by adding objects and conservatively lowering the target quantity. Thus, uncertainty is introduced into the instruction generation process in a controllable lower bound manner, which improves the reliability, robustness, and executability of scheduling allocation and reduces the risk of over- or under-scheduling.

[0111] A power load dispatching system based on the ARIMAX model and causal forest includes:

[0112] The data acquisition and preprocessing module is used to acquire scheduling modeling data of the target scheduling area, perform data processing, and generate preprocessed scheduling sample sequences.

[0113] The ARIMAX baseline prediction module is used to receive preprocessed scheduling sample sequences and train the ARIMAX baseline load prediction model, predict the target scheduling period, and output the baseline load prediction sequence.

[0114] The residual incremental construction module is used to receive the preprocessed scheduling sample sequence and the baseline load prediction sequence, subtract them at each time step to generate the incremental load sample sequence, and merge them to generate the causal training dataset.

[0115] The Honest Causal Forest Training Module is used to receive the causal training dataset and divide it into a structure learning subset and an effect estimation subset according to time blocks, generate a tree structure and train the Honest Causal Forest Model.

[0116] The effect inference and interval generation module is used to receive the honest causal forest model, locate the leaf nodes and read the leaf node effects, and aggregate the output of the processed effect estimation sequence and the processed effect confidence interval.

[0117] The fusion prediction and scheduling instruction generation module is used to perform effect fusion to obtain the policy condition load prediction sequence, filter the execution objects, calculate the allocation amount, and generate a scheduling instruction set.

[0118] Example 1: To verify the feasibility of the present invention in practice, it was applied to a comprehensive load dispatching scenario covered by a power distribution network. This scenario includes industrial loads, commercial buildings, public facilities, and user-side energy storage resources. The dispatching center needs to predict short-cycle loads under a unified sampling interval and issue dispatching actions such as demand response reduction, interruptible load limiting, and energy storage charging and discharging to achieve peak shaving, peak shifting, and safety margin control. The prominent problem faced by existing solutions in this type of scenario is that although load forecasting can provide a baseline trend, the actual reduction effect of dispatching actions is highly unstable under different objects, different operating conditions, and different external conditions. Using empirical coefficients or point estimations can easily lead to over-dispatch or under-dispatch, resulting in target deviation, execution failure, and difficulty in timely convergence of operational risks. Therefore, a dispatching method that can explicitly quantify the uncertainty of the effect of dispatching actions and use it for constraint allocation is needed.

[0119] In this scenario, the dispatch center continuously collects dispatch modeling data at fixed sampling intervals to form a preprocessed dispatch sample sequence. The sample records, indexed by the sampling time, synchronously include load observations, exogenous variables, dispatch actions, and object profile information. Exogenous variables originate from meteorological and market-side data streams and are aligned with the sampling time. The object profile consists of fields such as historical response capability, available time periods, and capacity boundaries. During the preprocessing stage, timestamp unification and granularity alignment are achieved. Missing segments are imputed using neighborhood interpolation within the same time period, and abnormal spikes are truncated and replaced with sliding windows. Simultaneously, time period features are constructed to enhance periodicity, resulting in a continuous, gap-free preprocessed dispatch sample sequence. In the baseline prediction stage, the preprocessed dispatch sample sequence is used to train the ARIMAX baseline load prediction model, and the baseline load prediction sequence is obtained by recursively extrapolating the target dispatch cycle. Subsequently, the difference between the load observations and the baseline load predictions is calculated time-by-time to obtain an incremental load sample sequence. The incremental load samples are then merged with the corresponding dispatch actions, exogenous variables, and object profiles to form a causal training dataset. To avoid temporal leakage and adaptive bias, the causal training dataset is divided into time blocks and used for tree structure learning and effect estimation respectively to generate an honest causal forest model. In the online inference stage, the preprocessed scheduling sample sequence of the target scheduling cycle is input into the model, the leaf nodes are located tree by tree and the leaf node effects are read. The drift of small sample leaf nodes is suppressed by effective sample size and stability screening and parent node backtracking. Then, the treatment effect estimation sequence is obtained by truncated weighted aggregation, and the confidence interval of the treatment effect is output at the same time.

[0120] The fusion phase integrates the processing effect estimation sequence with the baseline load prediction sequence to obtain the policy-conditional load prediction sequence. Confidence weights are generated using the confidence interval width to scale and limit the effect, preventing unstable effects at a few moments from skewing the load prediction. In the scheduling phase, the lower bound of the processing effect confidence interval is extracted to form a lower bound sequence. This lower bound sequence is then filtered using median filtering and backoff replacement to obtain a conservative lower bound sequence. At each sampling moment, candidate objects are screened based on conservative thresholds and sorted hierarchically according to conservative contribution density. Priority is given to calculating and limiting the allocation amount within the main execution layer, combining the object's executable upper and lower limits. The remaining target amount is updated in a closed loop according to the lower bound to ensure contribution. When accessibility is insufficient, a supplementary layer is introduced to add objects and repair the issue. If still insufficient, the target amount at that moment is reduced by a conservative proportion to ensure instruction execution. Finally, a scheduling instruction set containing object identifiers, sampling times, action types, and allocation amounts is generated and issued for execution, thus truly implementing "uncertainty" in scheduling allocation in the form of a controllable lower bound.

[0121] To verify the beneficial effects, three sets of comparative schemes were constructed under the same load scenario and the same sampling interval, and playback and joint testing were performed over multiple consecutive scheduling cycles: Comparative scheme 1 was ARIMAX baseline prediction + empirical reduction coefficient allocation; Comparative scheme 2 was ARIMAX baseline prediction + ordinary causal forest point estimation allocation; and the scheme of this invention was ARIMAX baseline prediction + honest causal forest effect interval + lower bound constraint allocation. Specific comparative data are shown in Table 1.

[0122] Table 1. Comparison of the effects of different scheduling strategies under comprehensive load scenarios.

[0123] index Comparison 1 Comparison 2 This invention Baseline Prediction 6.8% 6.7% 6.6% Peak reduction achievement rate 81.3% 86.9% 93.8% Instruction Execution Rate 88.5% 91.2% 97.1% Insufficient scheduling rate 12.4% 8.7% 3.1% Over-scheduling rate 9.6% 6.5% 2.4% Post-implementation load rebound rate 7.9% 5.8% 2.0% Number of times exceeding the limit (in the statistics window) 9 6 2 Comprehensive scheduling cost index (relative value) 1.00 0.93 0.86 index Comparison 1 Comparison 2 This invention Baseline prediction of MAPE (full cycle) 6.8% 6.7% 6.6% Peak reduction achievement rate 81.3% 86.9% 93.8% Instruction Execution Rate 88.5% 91.2% 97.1% Insufficient scheduling rate 12.4% 8.7% 3.1% Over-scheduling rate 9.6% 6.5% 2.4% Post-implementation load rebound rate 7.9% 5.8% 2.0% Number of times exceeding the limit (in the statistics window) 9 6 2 Comprehensive scheduling cost index (relative value) 1.00 0.93 0.86

[0124] As shown in Table 1, under the premise that the baseline prediction errors are basically the same (MAPEs of 6.8%, 6.7%, and 6.6%, respectively), the proposed scheme of this invention, which introduces the "honest causal forest effect interval + lower bound constraint allocation," demonstrates significant advantages in scheduling execution: the peak reduction achievement rate increases to 93.8%, higher than the 81.3% of the empirical coefficient allocation and the 86.9% of the point estimation allocation; the instruction executability rate reaches 97.1%, an increase of 8.6 percentage points and 5.9 percentage points compared to Control 1 and Control 2, respectively; and the underscheduling rate and The over-scheduling rates were reduced to 3.1% and 2.4%, respectively, significantly lower than 12.4% and 9.6% of Control 1 and 8.7% and 6.5% of Control 2. Meanwhile, the post-execution load rebound rate was reduced to 2.0%, and the number of overruns was reduced to 2, both significantly better than Control 1 (7.9%, 9 times) and Control 2 (5.8%, 6 times). On the cost side, the comprehensive scheduling cost index of this invention was 0.86, lower than 1.00 of Control 1 and 0.93 of Control 2, indicating that better comprehensive cost performance was achieved while improving achievement and safety.

[0125] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A power load dispatching method based on the ARIMAX model and causal forest, characterized in that, Includes the following steps: Collect scheduling modeling data for the target scheduling area, complete data preprocessing, and generate a preprocessed scheduling sample sequence; The ARIMAX baseline load prediction model is trained using preprocessed scheduling sample sequences to predict the target scheduling period, thus obtaining the baseline load prediction sequence for the target scheduling period. Based on the preprocessed scheduling sample sequence and the baseline load prediction sequence, residual incremental target construction is performed, incremental load sample sequence is calculated, and causal training dataset is generated by merging them. Honest training is performed on the causal training dataset by time-series block division, which is divided into a structure learning subset and an effect estimation subset. The structure learning subset is used to generate a tree structure, and the effect estimation subset is used for training to obtain an honest causal forest model. Input the preprocessed scheduling sample sequence corresponding to the target scheduling period into the honest causal forest model, perform honest causal forest effect inference and interval generation, and output the treatment effect estimation sequence and treatment effect confidence interval; A fusion operation is performed based on the treatment effect estimation sequence and the baseline load forecast sequence to obtain the policy-conditional load forecast sequence; The lower bound constraint scheduling of the effect interval is applied to the load prediction sequence and the confidence interval of the treatment effect. The scheduling action execution objects are selected based on the lower bound, the scheduling action allocation is calculated, and a scheduling instruction set is generated.

2. The power load dispatching method based on the ARIMAX model and causal forest according to claim 1, characterized in that, The specific steps to obtain the preprocessed scheduling sample sequence include: The scheduling modeling data of the power system in the target scheduling area is collected, and the scheduling modeling data is organized into a set of records arranged by sampling time according to a unified data recording format to obtain the original record sequence. Each record in the record set contains the load observation value, exogenous variables, scheduling actions and user or equipment profile information corresponding to the sampling time and the current sampling time. Data preprocessing is performed on the original record sequence, including timestamp unification, time granularity alignment, missing value imputation, outlier handling, and time period feature construction, and a preprocessed scheduling sample sequence is generated.

3. The power load dispatching method based on the ARIMAX model and causal forest according to claim 1, characterized in that, The acquisition of the baseline load forecast sequence for the target scheduling period specifically includes: Load observations are extracted from the preprocessed scheduling sample sequence in the order of sampling time to form a load time series, and exogenous variables corresponding to each sampling time are extracted simultaneously to form an exogenous variable sequence, and an aligned data table of sampling time-load observation-exogenous variable is established. Differential processing is performed on the load time series. The difference between the multi-order historical load observations is calculated according to the preset differential order to obtain the current differential value. The final differential values ​​corresponding to each sampling time are arranged in chronological order to obtain the differential load series. In the aligned data table, the differential load series replaces the original load time series as the modeling sequence, while retaining the end historical load observations before differential processing. An ARIMAX model is built and trained on the differential load sequence. Candidate combinations of autoregressive and moving average orders are set, and corresponding model input structures are constructed based on exogenous variables in the preprocessed scheduling sample sequence. Parameter estimation is performed on each candidate order combination, the corresponding model fitting error is calculated, and the order combination with the smallest fitting error is selected as the final model structure. After determining the model structure, parameter estimation is performed on the autoregressive coefficients, moving average coefficients, and exogenous variable coefficients, and model training is completed. The exogenous variable sequence corresponding to the target scheduling period is input into the trained ARIMAX model, and the differential prediction sequence of the target scheduling period is generated recursively in time order. Combined with the retained historical load observations at the front and rear of the differential sequence, inverse differential reconstruction is performed on the differential prediction sequence time by time to obtain the baseline load prediction sequence of the target scheduling period.

4. The power load dispatching method based on the ARIMAX model and causal forest according to claim 1, characterized in that, The specific steps involved in obtaining the causal training dataset are as follows: Extract data segments from the preprocessed scheduling sample sequence that cover the same target scheduling period as the baseline load prediction sequence, and align them with the baseline load prediction sequence time by time to form an aligned record sequence; The residualized incremental target is constructed time-by-time for the aligned record sequence. The load observation value at each sampling time is subtracted from the baseline load prediction value at the same sampling time to obtain the incremental load sample value at the current sampling time. The incremental load sample sequence is obtained by arranging them in chronological order. The incremental load sample sequence is merged with the exogenous variables, scheduling actions, and user or device profiles in the aligned record sequence according to the sampling time to generate a causal training dataset.

5. The power load dispatching method based on the ARIMAX model and causal forest according to claim 1, characterized in that, The honest causal forest model is obtained specifically through the following steps: The causal training dataset is divided into time-series blocks according to the sampling time sequence. The block length is set to the preset number of sampling points. Starting from the initial sampling time, the dataset is divided into continuous time blocks according to the block length, and each time block is assigned a block number. Honest partitioning is performed on each time block. A subset of time blocks are selected to form a structure learning subset, and the remaining time blocks that do not overlap with the structure learning subset are selected to form an effect estimation subset. The set of block numbers corresponding to the structure learning subset and the effect estimation subset is recorded. By using the structure learning subset to generate tree structure, sample subsampling is performed on each tree. Candidate splitting features are selected from the covariate field within each node to be split, and candidate splitting thresholds are traversed. For each candidate splitting threshold, the node sample is divided into left and right child nodes. At the same time, the splitting score is calculated based on the difference in the processing effect of the left and right child nodes. The candidate splitting feature with the highest splitting score and the candidate splitting threshold are selected. Repeat the node splitting process for each tree until the preset stopping condition is met, and obtain the tree structure set determined by the structure learning subset; Leaf node effects are estimated using a subset of effect estimates. Samples in the subset of effect estimates are assigned to corresponding leaf nodes according to the tree structure path. Within each leaf node, the mean of the result field of the treated state sample and the mean of the result field of the untreated state sample are calculated and subtracted to obtain the estimated value of the treatment effect of the leaf node. The estimated values ​​of the treatment effect of the leaf nodes of each tree are encapsulated with the tree structure set to form an honest causal forest model.

6. The power load dispatching method based on the ARIMAX model and causal forest according to claim 1, characterized in that, The outputs of the treatment effect estimation sequence and the treatment effect confidence interval specifically include: Read the preprocessed scheduling sample sequence corresponding to the target scheduling period, extract the covariate field and scheduling action field of each sampling time in the order of sampling time, construct the model input sample of the sampling time, and write the model input sample into the inference queue; The model input samples are sequentially retrieved from the inference queue and input into each tree of the honest causal forest model. The splitting features and splitting thresholds of the current node are read in each tree. Layer-by-layer comparison and branch selection are completed until the leaf node is reached. The leaf node identifier on each tree at the sampling time is recorded, and the effective sample quantity identifier and leaf node stability identifier of the leaf node are read simultaneously. For each tree, the leaf node output leaf node identifier is filtered for validity. Leaf node results with a valid sample size less than a preset minimum threshold or stability lower than a preset threshold are removed. The removed leaf node results are backtracked and replaced with their parent nodes. After the filtering and backtracking are completed, the pre-stored leaf node processing effect estimate value of the corresponding leaf node is read to form a cross-tree processing effect estimate value set. Weighted robust aggregation is performed on the cross-tree treatment effect estimate set. For each treatment effect estimate in the set, the corresponding effective sample size identifier is read as the weight. The treatment effect estimate set is truncated according to the weight to remove extreme values ​​of a preset proportion. At the same time, the remaining treatment effect estimates are weighted and averaged to obtain the treatment effect estimate at the current sampling time. The results are then arranged in chronological order to generate the treatment effect estimate sequence. Interval generation is performed on the set of cross-tree treatment effect estimates. The treatment effect estimates in the set are sorted by weight. The lower and upper quantile values ​​corresponding to the preset confidence level are extracted to obtain the original confidence interval. At the same time, consistency shrinkage is performed on the original confidence interval to remove trees that deviate from the treatment effect estimates by more than a preset threshold. The upper and lower bounds are then output and recalculated. The confidence interval will be subjected to time consistency correction to eliminate abrupt changes between adjacent time intervals, and the confidence interval of the processing effect will be output.

7. The power load dispatching method based on the ARIMAX model and causal forest according to claim 1, characterized in that, The specific steps to obtain the strategy conditional load prediction sequence include: Read the baseline load prediction sequence, treatment effect estimation sequence and treatment effect confidence interval of the target scheduling period, check the time index at each sampling time and complete the filling and elimination, and calculate the confidence weight corresponding to the width of the treatment effect confidence interval for each sampling time. The target scheduling cycle is traversed in the order of sampling time. The estimated value of the treatment effect is scaled according to the confidence weight to obtain the corrected treatment effect value. At the same time, the corrected treatment effect value is restricted to the upper and lower bounds of the confidence interval of the treatment effect at the current sampling time to obtain the amplitude-limited treatment effect value, which is added to the baseline load forecast value to obtain the policy condition load forecast value. The policy condition load prediction values ​​obtained at each sampling time are written into the result sequence in chronological order to form the policy condition load prediction sequence.

8. The power load dispatching method based on the ARIMAX model and causal forest according to claim 1, characterized in that, The generation of the scheduling instruction set specifically includes: Read the policy condition load prediction sequence and the confidence interval of the treatment effect. Extract the lower bound of the confidence interval at each sampling time to form a lower bound sequence. Perform median filtering on the lower bound sequence to suppress abrupt change points. Perform back-down replacement on the lower bound value that shows a reverse jump in a continuous time period to obtain a conservative lower bound sequence. For each sampling time, the current lower bound value of the conservative lower bound sequence is read, and objects that have not reached the conservative threshold are removed to obtain a set of candidate objects. The conservative contribution density of each candidate object is calculated and sorted in descending order of conservative contribution density to form a sorted object list. The sorted object list is divided into the main execution layer and the supplementary layer. For each sampling time, starting from the main execution layer, read the lower limit and upper limit of executable allocation of each object, generate the initial allocation of objects based on the remaining scheduling target quantity, perform a limit on the initial allocation of objects to obtain the object allocation quantity, and write the object allocation quantity into the allocation result set. For each sampling time, the total lower bound guaranteed contribution is obtained by summarizing the allocation result set and comparing it with the scheduling target amount. When the total lower bound guaranteed contribution is insufficient, objects are added from the supplementary layer and the allocation amount calculation is repeated until the target amount is met or there are no available objects. The scheduling target amount is then reduced by a preset conservative ratio and the allocation result set is recalculated. The allocation results of each sampling time within the target scheduling period are summarized in chronological order to generate a scheduling instruction set.

9. A power load dispatching system based on the ARIMAX model and causal forest, executing the power load dispatching method based on the ARIMAX model and causal forest as described in any one of claims 1 to 8, characterized in that, include: The data acquisition and preprocessing module is used to acquire scheduling modeling data of the target scheduling area, perform data processing, and generate preprocessed scheduling sample sequences. The ARIMAX baseline prediction module is used to receive preprocessed scheduling sample sequences and train the ARIMAX baseline load prediction model, predict the target scheduling period, and output the baseline load prediction sequence. The residual incremental construction module is used to receive the preprocessed scheduling sample sequence and the baseline load prediction sequence, subtract them at each time step to generate the incremental load sample sequence, and merge them to generate the causal training dataset. The Honest Causal Forest Training Module is used to receive the causal training dataset and divide it into a structure learning subset and an effect estimation subset according to time blocks, generate a tree structure and train the Honest Causal Forest Model. The effect inference and interval generation module is used to receive the honest causal forest model, locate the leaf nodes and read the leaf node effects, and aggregate the output of the processed effect estimation sequence and the processed effect confidence interval. The fusion prediction and scheduling instruction generation module is used to perform effect fusion to obtain the policy condition load prediction sequence, filter the execution objects, calculate the allocation amount, and generate a scheduling instruction set.

Images