An energy station operation strategy rule extraction method, system and device based on a multi-level decision tree of causal logic and a storage medium

Abstract

The application discloses an energy station operation strategy rule extraction method and system based on a causal logic inference multi-level decision tree, a device and a storage medium, relates to the technical field of global collaborative optimization and regulation of comprehensive energy systems, and comprises the following steps: performing Granger causality test on time series data, determining one-way causal driving direction, constructing a Granger causality topology network, and obtaining a standardized causal logic hierarchical framework; splitting energy station control tasks into multiple single-output control actions, screening driving equipment operating states as features, and constructing a dedicated feature set; using an independent decision tree to perform single-target independent training, fitting the mapping relationship between the current control action and the corresponding feature set; extracting independent regulation rules of the corresponding control action, and using the standardized causal logic hierarchical framework as a guide to sort out the independent regulation rules and generate complete energy station global operation rules. The method disclosed by the application significantly improves the prediction accuracy and eliminates irrelevant noise interference.

Description

Technical Field

[0001] This invention relates to the field of integrated energy system global collaborative optimization and control technology, specifically to a method, system, device, and storage medium for extracting energy station operation strategy rules based on causal logic inference multi-level decision tree. Background Technology

[0002] With the coordinated evolution of large-scale integrated energy systems, modern regional energy stations typically integrate various types and specifications of energy supply and distribution equipment, such as chillers, energy storage tanks, heat pumps, and water pumps. Because these devices exhibit extremely complex linkage and coupling characteristics in energy transfer, hydraulic circulation, and temporal response, accurately extracting highly reliable and interpretable station-wide coordinated operation strategies and control rules from historical operating data has become the strategic core for ensuring the safe, stable, and economically efficient operation of modern energy stations.

[0003] However, existing data-driven rule extraction methods generally rely on a single, flat, traditional decision tree algorithm, directly using all environmental parameters, load characteristics, and equipment status of the energy station as full input. Figure 1 The joint control rules for the entire system are directly fitted step-by-step. However, this globally unified strategy generation paradigm suffers from a serious drawback: it lacks a rational examination of the temporal evolution relationships between control actions and the physical context of the "drive-response" relationship. Traditional full-feature input methods often lead to decision trees falling into data correlation traps when recursively dividing feature nodes, failing to identify the unidirectional causal driving direction between multiple interconnected control actions. This lack of causal logical inference causes the model to treat numerous spurious and accidental statistical correlations as control mechanisms when solving for the splitting threshold, resulting in a significant degradation in model decision accuracy under complex operating conditions.

[0004] Given this lack of clarity regarding causal topology, existing control decision-making fitting processes typically involve simultaneous fitting of multiple outputs, mixing discrete start-stop control with continuous setpoint control. This interweaving of multiple control actions significantly increases the fitting dimensionality of individual models and the computational complexity of finding the optimal splitting path. Furthermore, the lack of decoupling modeling for heterogeneous variables and fine-grained scale constraints leads to excessive redundancy or overfitting of rules for certain key operating conditions. Low-probability operating condition noise and abnormal fluctuations in historical operating data propagate layer by layer along the complex tree structure, directly resulting in a massive and cumbersome control rule structure with uncontrolled levels of judgment condition linkages. This lack of intuitiveness and concise interpretability poses a significant obstacle to on-site verification and deployment in actual engineering projects.

[0005] More seriously, when these multi-objective control rules, which are at different levels and have strict physical transmission sequences, are ultimately stacked together, existing technologies lack an effective systematic hierarchical architecture to provide temporal guidance and physical constraints for each independent sub-rule. Because continuous and discrete control rules inherently have a vertical hierarchical relationship in terms of time step and driving level based on HVAC physics, a flat rule system cannot form a dynamic cascaded control mechanism at the entire station scale. This not only leads to misalignment in the timing of the final generated operational strategy rules, but also fails to prevent logical conflicts between the action results of upstream driving equipment and the judgment conditions of subsequent response equipment. This frequently causes state oscillations or operational condition prediction deviations in the actual scheduling and matching of station-wide rules, seriously threatening the robustness and coordinated consistency of the overall operational rules of the regional energy station. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is that existing data-driven rule extraction methods suffer from problems such as missing causal logic, mixed heterogeneous control targets, high computational complexity, and lack of a systematic hierarchical architecture.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a method for extracting energy station operation strategy rules based on a multi-level decision tree using causal logic inference. This method includes performing Granger causality tests on the time-series data of control variables for each device in the energy station; determining the unidirectional causal driving direction between control variables based on the test results; constructing a Granger causal topology network; sorting the network topologically to determine the causal hierarchy; obtaining a standardized causal logic hierarchy framework; breaking down the energy station control task into multiple single-output control actions; selecting the operating status of the driving equipment at the preceding level of the current control action as features based on the standardized causal logic hierarchy framework; constructing a unique feature set for each single-output control action; for each single-output control action, performing independent training using independent decision trees, ensuring that each independent decision tree only fits the mapping relationship between the current control action and the corresponding feature set; extracting independent control rules for the corresponding control actions from the independent decision trees; and using the standardized causal logic hierarchy framework as a guide, summarizing and organizing the independent control rules hierarchically according to the causal transmission order of each control action within the framework to generate complete operation rules for the entire energy station.

[0009] As a preferred embodiment of the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree described in this invention, the following steps are taken: before performing Granger causality test, the time series data of each equipment control variable are tested for stationarity; for non-stationary time series data that fail the stationarity test, the time series trend is eliminated and transformed into a standardized time series that meets the preset stationarity conditions, which serves as the data source for causal determination.

[0010] As a preferred embodiment of the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree described in this invention, the construction of Granger causal topology network includes applying physical unidirectional hierarchical constraints during the causal verification process, limiting the generation of causal driving edges only from high-level control variables in HVAC physics to low-level control variables, or only allowing mutual influence between control variables at the same level, and using the physical unidirectional hierarchical constraints to correct the topology network generated by pure data driving.

[0011] As a preferred embodiment of the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree described in this invention, the method comprises: when constructing a dedicated feature set for a single output control action, selecting the operating status of the driving equipment located at the preceding level of the current single output control action in the standardized causal logic hierarchical framework as internal physical linkage features; introducing system load variables and time-of-use electricity price variables determined by the external environment and market mechanisms as external boundary driving features; and combining the internal physical linkage features with the external boundary driving features to construct a dedicated feature set for each single output control action.

[0012] As a preferred embodiment of the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree described in this invention, the independent decision tree is trained independently for a single objective, and the single output control action to be optimized is divided into discrete control action and continuous control action; for discrete control action, the corresponding independent decision tree is trained using a classification splitting criterion to output the control rules of discrete categories; for continuous control action, the corresponding independent decision tree is trained using a regression splitting criterion to output the continuous optimal setting value rules.

[0013] As a preferred embodiment of the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree described in this invention, the independent decision tree, when undergoing single-objective independent training, further includes using a pre-pruning constraint strategy to constrain the independent training growth process of the independent decision tree. The pre-pruning constraint strategy includes at least one of limiting the maximum depth of each decision tree, constraining the minimum sample capacity of leaf nodes, and setting a minimum impurity reduction amount. Limiting the maximum depth of each decision tree limits the maximum linkage level of the conditional judgment sentences in the extracted independent control rules. Constraining the minimum sample capacity of leaf nodes removes abnormal operating conditions with a probability lower than a preset threshold from historical data. Setting a minimum impurity reduction amount terminates tree growth early when the distinguishability of the precondition for the control action is lower than a preset gain, and merges control state branches with low contribution.

[0014] As a preferred embodiment of the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree described in this invention, the generation of complete operation rules for the entire energy station includes: traversing all decision paths from the root node to all leaf nodes in each independent decision tree, and converting the multi-dimensional threshold branch judgment logic in each decision path into standardized condition expression rules; during hierarchical summary and sorting, the execution results of the condition expressions generated in the previous level are dynamically cascaded as the condition judgment input for triggering the condition expressions in the subsequent response level, thereby realizing global cascaded control.

[0015] Another objective of this invention is to provide an energy station operation strategy rule extraction system based on a causal logic inference multi-level decision tree. This system can break down the energy station control task into multiple single-output control actions, and based on a standardized causal logic hierarchical framework, select the operating status of the driving equipment located at the previous level of the current control action as features to construct a unique feature set for each single-output control action. This solves the problem of heterogeneous control targets in current data-driven rule extraction methods.

[0016] As a preferred embodiment of the energy station operation strategy rule extraction system based on causal logic inference multi-level decision tree described in this invention, it includes: a time-series data preprocessing and causal topology inference module, a multi-objective action decoupling and dedicated feature configuration module, a classification regression calculation array and scale pre-pruning module, and a global rule expression dynamic cascading integration module; the time-series data preprocessing and causal topology inference module is used to collect time-series operation data of energy station equipment, perform stabilization differential correction and causal verification, and inject unidirectional physical hierarchical constraints based on HVAC mechanisms to construct a standardized causal logic hierarchical framework; the multi-objective action decoupling and dedicated feature configuration module is used to decompose the linkage control task into multiple independent single-output control actions according to the physical action class, and, based on the standardized causal logic hierarchical framework, limit the selection to only external boundary features and driving equipment variables at the preceding level of the current control action, and configure each single-output control action... The system constructs a dedicated feature set; the classification and regression computation array and scale pre-pruning module are used to allocate independent decision trees to independently train the dedicated feature sets of each single-output control action, and to divide the action according to the action type. For discrete actions, a classification splitting criterion is used, and for continuous actions, a regression splitting criterion is used. A pre-pruning strategy is introduced, which controls the number of condition judgment levels, removes operating condition noise, and merges low-discrimination branches by limiting the tree depth, constraining the sample size of leaf nodes, and setting a minimum impurity reduction amount. The global rule expression dynamic cascading integration module is used to traverse the decision paths of each decision tree in the computation array and uniformly transform them into standardized condition expression rules. Following the vertical transmission chain in the hierarchical framework, the execution result of the preceding driving hierarchical expression is dynamically used as the condition judgment input of the subsequent response hierarchical expression. Finally, it is integrated into a global complete operation rule that satisfies the HVAC physical coupling mechanism and eliminates temporal action conflicts.

[0017] Another objective of this invention is to provide an energy station operation strategy rule extraction device based on a causal logic inference multi-level decision tree, comprising a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the energy station operation strategy rule extraction method based on a causal logic inference multi-level decision tree.

[0018] Another object of the present invention is to provide a storage medium for extracting energy station operation strategy rules based on a causal logic inference multi-level decision tree, wherein a computer program is stored thereon, and when the computer program is executed by a processor, it implements the steps of the method for extracting energy station operation strategy rules based on a causal logic inference multi-level decision tree.

[0019] The beneficial effects of this invention are: The energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree provided by this invention introduces a pre-emptive Granger causality test in the feature engineering stage, injecting unidirectional physical hierarchical constraints based on HVAC mechanisms. At both mathematical and engineering mechanism levels, it eliminates pseudo-causal and accidental statistical correlation features caused by strong coupling of multiple devices, accurately capturing the true physical boundary points of heterogeneous devices under critical transition conditions. By splitting the linkage control task into independent single-output control actions according to physical action categories and performing single-objective focused training, it eliminates the multidimensional variable cross-interference during multi-objective synchronous fitting. The method also introduces pre-pruning constraint strategies such as limiting the maximum tree depth and constraining the minimum sample capacity of leaf nodes, effectively filtering extreme probability noise and restricting conditional expressions. The maximum linkage level transforms complex mappings into standard conditional expressions of "If-condition-Then-action," resulting in a streamlined and non-redundant rule set for each device, significantly improving readability and verification convenience during practical engineering implementation. Using a standardized causal logic hierarchy framework as the core integration guide, and following the physical coupling mechanism of HVAC systems, rules are hierarchically summarized and cascaded. This allows the execution results of preceding driving hierarchical expressions to dynamically cascade as conditional inputs for triggering subsequent response hierarchical expressions, establishing a vertical "cause-effect" transmission chain at the entire station scale. This completely severs the temporal action conflicts when synchronously fitting continuous and discrete control targets, ensuring extremely high operating condition prediction accuracy while endowing the energy station's overall operation rules with unparalleled physical consistency and engineering robustness. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is an overall flowchart of the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree in this invention.

[0022] Figure 2 This is a Granger causal topology network diagram for the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree in this invention.

[0023] Figure 3 This is a schematic diagram of the overall system for extracting energy station operation strategy rules based on causal logic inference multi-level decision trees, which is the method for extracting energy station operation strategy rules based on causal logic inference multi-level decision trees in this invention. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0027] Example 1

[0028] Reference Figure 1 This is the first embodiment of the present invention, which provides a method for extracting energy station operation strategy rules based on a multi-level decision tree inference using causal logic.

[0029] Specifically, a causal hierarchy is constructed for the time-series data of each equipment control variable. During the data preparation phase, multi-dimensional time-series operational data, including total load, time-of-use electricity price, various unit statuses, and pump parameters, are collected using a uniform sampling step of one hour. This data is then time-aligned and stabilized to obtain a standardized stationary sequence dataset. Subsequently, the Granger causality test (101 method) is used to perform pairwise autoregressive analysis and determination on the extracted equipment control variable data. In the determination process, variable A is considered a Granger cause of variable B only if the introduction of historical time-series data of variable A significantly improves the prediction accuracy of the future operating conditions of variable B. A significance threshold is set (commonly α=0.05), and the significance is determined by calculating the corresponding p-value. If the p-value is less than 0.05, a unidirectional causal relationship from the driving variable to the response variable is determined. This is achieved by traversing the pairwise autoregressive analysis of all equipment control variables. Two sets of data were combined, and all non-causal pseudo-correlation features with p-values ​​greater than 0.05 were removed, retaining only strong causal feature variables with significant time-driven relationships. Finally, using the control variables of each device as nodes and the determined Granger causal directions as directed edges, a Granger causal topology network graph 102 was constructed. Introducing professional knowledge from the HVAC field and the characteristics of on-site operation, unidirectional hierarchical physical constraints were applied to the topology to limit the influence of higher-level variables on lower-level variables. The topology graph was then sorted to determine the causal sequence of each control action, thus obtaining a standardized causal logical hierarchy framework. In this framework, centrifugal chillers and magnetic levitation chillers, as basic cooling equipment, are at the first level; dual-condition chillers and heat pumps, as load peak-shaving equipment, are at the second level; the opening and closing of water tank valves and the number of cold storage devices, as devices determining the energy supply / release mode, are at the third level; the outlet water temperature, as a characterization of the heat exchange state of the terminal system, is at the fourth level; and the pump frequency, as a pure responder, is at the fifth level.

[0030] Furthermore, the decoupling and decomposition of control actions and the construction of a dedicated feature set 202 are carried out. This decouples the originally complex, multi-input, multi-output coupled energy regulation system into independent single-output modeling sub-tasks for each control device and each regulation target. When constructing a subset of training samples for each independent control action, the corresponding historical time-series values ​​are extracted as the model's output labels. Based on the defined causal logic hierarchy framework, the system strictly constrains and filters the feature selection range according to causal logic: for the single-output control action 201 that needs to be modeled, only the operating status of equipment or environmental variables (such as external system load, time-of-use electricity price, start-up or shutdown of upper-level driving equipment, etc.) located at its preceding level in the hierarchy framework are selected as input features. This allows for the customized construction of a subset of strong causal feature inputs for each control action, free from redundant noise interference. For example, for the opening and closing action of a water tank valve at the third level, its exclusive feature set 202 only retains the number of first-level chillers supplying cooling and the operating status of the same-level cold storage equipment that have a significant driving effect on it.

[0031] Furthermore, a decision tree model is trained for each independent control action. Independent decision tree 301 (CART decision tree) is used to train each decoupled single-output subtask independently for a single objective. This ensures that each independent decision tree 301 focuses solely on fitting the mapping relationship between the current single control objective and its corresponding exclusive strong causal feature subset during its growth, without needing to consider the control effects of other devices. The model distinguishes based on the variable type of the control action: if the control action is discrete (such as the number of devices started / stopped, valve switching status, etc.), a CART classification tree is constructed and trained, using the Gini coefficient as the node splitting parameter. The criteria ultimately output discrete category control rules in the form of "IF-THEN". If the control action is a continuous action (such as the pump frequency setpoint, the unit outlet water temperature, etc.), a CART regression tree is constructed and trained. The minimum mean square error (MSE) is used as the node splitting criterion, and the rules corresponding to the continuous optimal setpoint values ​​are output. In order to avoid the risk of overfitting from a single decision tree and to ensure the intuitive interpretability of the rules, a pre-pruning strategy is uniformly adopted during the training process. By strictly limiting the maximum depth of each tree, the minimum sample size of the leaf nodes, and the minimum decrease in impurity, the decision tree is controlled within a shallow tree structure and a simple decision path.

[0032] The extraction and hierarchical integration of the complete operation strategy rules for the entire station were carried out. After the CART decision trees for each single objective were trained independently, all decision paths from the root node to the leaf node of each tree were traversed. The multi-dimensional splitting features and splitting threshold judgment logic on each path were automatically transformed into standardized "If-condition-Then-action" rules. This generated a set of concise, non-redundant, and independent control rules corresponding to the physical mechanism for each device. The standardized causal logic hierarchical framework 103 was used as the core integration basis. The physical causal transmission sequence from the first level to the fifth level was strictly followed. The independent control rules of all single devices were summarized, sorted, and cascaded in a hierarchical manner. Since the entire integration process was guided by the causal logic structure, the rule branch interference caused by the synchronous fitting of continuous and discrete control objectives was completely eliminated. This avoided problems such as control rule matching errors, operating condition prediction deviations, or rule timing conflicts. Finally, a set of complete operation rules 401 for the entire energy station was successfully generated, which takes into account physical causal consistency, high operating condition matching prediction accuracy, clear logic, and high interpretability.

[0033] Example 2

[0034] Reference Figure 1 and Figure 2 This is the second embodiment of the present invention. Unlike the previous embodiment, this embodiment provides a stationary causal topology network construction that incorporates unidirectional constraints of HVAC engineering mechanisms. This solves the problem that traditional extraction methods have a long-term linear drift trend with non-stationarity and are prone to data correlation traps and spurious correlation interference in autoregressive solutions.

[0035] Specifically, stationarity tests and trend elimination are performed on the time-series data of control variables for each device, multi-dimensional time-series data of regional energy stations are collected, and original multivariate time-series datasets are constructed. An ADF unit root stationarity test is performed on all variable sequences within the dataset, specifically as follows: , in, The dependent variable is the predictive target of ADF regression. The constant intercept term is the regression fitting constant, which is an arbitrary real number. The core test coefficient of the unit root. It has a unit root and is non-stationary. No unit root, stationary. This is a term with a lag of 1 in the original sequence. It is a linear time trend coefficient. This indicates no long-term linear trend. Including trends, These are the original observations from the previous time step. This represents the maximum order of the lag difference in the ADF test. For the first The regression weight coefficients corresponding to the first lag difference, the lag of the difference sequence Time value, For lag The first-order difference term of the order, the difference sequence lags. Time value, The noise residuals follow a normal white noise with a mean of 0, and there is no fixed value; the residuals are different at each time step.

[0036] If the ADF test p-value is less than the preset stationarity threshold of 0.05, the series is determined to be a stationary series. For non-stationary time series data that fail the stationarity test and whose p-value is greater than or equal to the preset stationarity threshold of 0.05, the system uses first-order differencing to eliminate long-term or local linear trends in the time series, transforming it into a standardized stationary time series that meets the preset stationarity conditions, specifically as follows: , in, for Original time-series observations at any given time, and actual measured values ​​from the equipment for the current hour. for For non-stationary sequences, the original time series observations at time points are used... The Granger causality test 101 is performed as input. The final output is a standardized time series dataset that fully satisfies the stationarity requirement, providing input samples for subsequent Granger causality modeling.

[0037] Based on a stationary data source, a Granger causal topology network 102 with physical unidirectional hierarchical constraints is constructed. For the obtained stationary time-series variables, a bivariate VAR autoregressive model is constructed. A regression equation with lag order k is established for any two device time-series variables X and Y, expressed as: , , in, for Time variable stationary time series observations, For the timing of the upper-layer driver devices; for Time variable stationary time series observations, Timing for lower-level response devices; Current time-series index, a positive integer ; For cyclic indexes of lag order, , representing lag One time step; The maximum lag order of the VAR model; for Value of a stationary sequence with lag order i; for Lag The value of a stationary sequence of order 1; for Regression Lag Order coefficient, a measure of history For the present Self-driving strength; for right The lagged cross-regression coefficient measures historical data. For the present Influence weight; for right The lagged cross-regression coefficient measures historical data. For the present Influence weight; for Regression Lag Order coefficient, a measure of history For the present Self-driving strength; for The random disturbance term (residual) of the equation is white noise with a mean of 0, which is independent at each time step and has no fixed value. for The equation's random disturbance term (residual) is white noise with a mean of 0, independent at each time step, and has no fixed value.

[0038] In constructing a directed causal topology network, to address potential issues like causal inversion or inconsistencies with engineering realities that might arise from purely data-driven approaches, a strong unidirectional hierarchical physical constraint is actively applied to the decision chain of verification and topology construction. This constraint limits the influence of control variables at higher levels in HVAC physics to those at lower levels, and allows mutual influence between control variables at the same level. In the specific verification process, all equipment variables are combined pairwise, and an F-test is performed based on the VAR model at a 95% confidence level. Only when... A significant, unidirectional causal driving direction is determined to exist only when the overall lag coefficient is significantly non-zero (i.e., the calculated significance p-value is less than 0.05) and the direction of influence does not violate the unidirectional hierarchical physical constraint. If the p-value is greater than 0.05, it is judged as a spurious correlation feature without causality and is removed. Figure 2As shown, through this dual constraint, a Granger causal topology network (Figure 102) is finally established, with each device's control variables as nodes and the verified Granger causal directions as red directed arrows. Due to the addition of hierarchical constraints, the centrifugal chiller and magnetic levitation chiller in the first level serve as the basic cooling hosts of the energy station, exhibiting extremely high driving capabilities and frequencies for multiple devices throughout the station; the dual-condition chiller and heat pump in the second level mainly undertake peak-shaving functions; the opening and closing of water tank valves and the number of cold storage devices in the third level determine the cooling supply and release mode; the outlet water temperature in the fourth level characterizes the heat exchange result status of the system; and the pump group frequency variable in the fifth level at the very bottom exists only as a responder and effect variable, without any driving logic output. By performing topological sorting on this causal topology graph that conforms to the physical mechanism, the causal sequence hierarchy of each control action can be divided, and a standardized causal logic hierarchy framework can be obtained. This completely eliminates irrelevant noise from both the data source and physical mechanism perspectives, thus removing interference from the subsequent multi-level independent decision tree (Figure 301) feature splitting.

[0039] Example 3

[0040] Reference Figure 1 This is the third embodiment of the present invention. Unlike the previous embodiment, this embodiment provides a multi-objective action decoupling and refined pre-pruning strategy to solve the problem that the existing control decision process intertwines multiple types of control actions such as discrete start-stop and continuous setpoints for multi-output fitting, which leads to increased computational complexity in finding the optimal split path, bloated and large extracted rule structure, and uncontrolled linkage levels of condition judgment sentences.

[0041] Specifically, regarding the construction process of the dedicated feature set for a single-output control action 201, the dedicated feature set 202 is composed of internal physical linkage features and external boundary driving features. The selection of internal physical linkage features is strictly constrained by the constructed standardized causal logic hierarchical framework 103. The system traverses and identifies all driving devices located at the preceding level of the current single-output control action 201 in the hierarchical framework, extracts their historical operating states as internal physical linkage features, and uses them to qualitatively describe the real causal linkage relationship between multiple coupled devices. For example, for the opening and closing action of the water tank valve at the third level, its dedicated feature set 202 is as follows: The feature set 202 retains the number of first-level chillers that significantly drive the operation and the operating status of the same-level cold storage equipment. The external boundary driving features are constructed by introducing system load variables and time-of-use electricity price variables determined by the external environment and market mechanisms to capture the comprehensive impact of external energy demand and economic boundaries on energy station strategies. Finally, the system organically combines the selected internal physical linkage features with the external boundary driving features to construct a unique and complete feature subset for each decoupled single-output control action 201, thereby limiting the range of feature selection by using causal logic and boundary constraints.

[0042] For each single-output control action 201, a classification or regression decision tree model is trained based on the constructed feature set. All single-output control actions 201 requiring optimization across the entire station are rigorously defined and their variable types are categorized. Discrete control actions specifically include the number of equipment starts / stops and the switching of water tank valve modes. For these discrete control actions, the system uses a CART classification tree for independent training on a single target, with the Gini coefficient as the internal splitting criterion and the class label as the output. Continuous control actions, such as pump frequency (Hz) and unit outlet water temperature (°C), are trained using a CART regression tree, with the minimum mean square error (MSE) as the splitting criterion and the leaf node mean as the output. Pre-pruning (limiting maximum depth, minimum number of leaf node samples, and minimum reduction in impurity) is used to prevent overfitting. The Gini coefficient and minimum mean square error are expressed as follows: , , in, For the sample set, For the first The proportion of samples of each class For the sample The continuous target value, For nodes Internal target.

[0043] The decision tree traverses all causal features and thresholds at each level, selecting the feature and threshold with the largest decrease in splitting exponent to complete node splitting, and growing layer by layer to obtain a single-device-specific decision tree model. To avoid overfitting of a single tree, this invention employs a pre-pruning method (limiting the maximum depth and the minimum number of leaf node samples) to prevent overfitting.

[0044] After model training, all decision paths from the root node to all leaf nodes are traversed. The multi-dimensional threshold judgment logic of each path is transformed into a standardized "If-condition-Then-action" control rule. Prediction is performed on each independent decision tree 301, and performance indicators are calculated: for discrete actions, the confusion matrix is ​​calculated; for continuous actions, RMSE, MAE, and R² are calculated. Each device independently generates a simplified, non-redundant sub-rule set corresponding to the physical mechanism, replacing the traditional single-tree full-scale complex rule system.

[0045] Finally, to ensure the generated rules are highly interpretable, concise, and free of redundancy, and to completely suppress the risk of model overfitting, a pre-pruning constraint strategy is uniformly introduced to strictly constrain them during the single-objective independent training and growth process. This pre-pruning constraint strategy includes at least one of the following: limiting the maximum depth of each decision tree, constraining the minimum sample capacity of leaf nodes, and setting a minimum impurity reduction. By limiting the maximum depth of each decision tree, the system can limit the maximum linkage level of conditional statements (i.e., the number of feature combinations in IF statements) in the final extracted independent control rules from the source, thereby controlling the tree structure to a shallow depth and preventing the generation of overly redundant and complex logical chains. By constraining the minimum sample capacity of leaf nodes… Load capacity (i.e., the minimum number of samples allowed in a leaf node) can forcibly filter and remove abnormal operating condition branches that have a very low probability of occurrence in historical operating data and are caused by niche extreme operating conditions or data noise, ensuring the universality of the rules. By setting a minimum impurity reduction amount (i.e., the minimum reduction amount of the Gini coefficient or mean square error when a node splits), the downward growth of the tree can be automatically terminated in advance when the distinguishability of the current control action by the preconditions is lower than the preset gain threshold. This allows control state branches that contribute less to the model's control accuracy to be merged nearby, thereby greatly simplifying the rule size of the tree and ensuring that the final generated global energy station operation strategy rules are well-organized and easy for engineers to read and understand intuitively.

[0046] Example 4

[0047] Reference Figure 1 This is the fourth embodiment of the present invention. Unlike the previous embodiment, this embodiment provides standardized condition expression rules based on a hierarchical topology framework. This solves the problem that a flat rule system cannot form a dynamic cascading control mechanism at the whole-site scale, which in turn causes logical conflicts or timing misalignments between the action results of the front-end drive equipment and the judgment conditions of the subsequent response equipment, resulting in state oscillations and operational condition prediction deviations during matching and scheduling.

[0048] Specifically, the system automatically traverses all decision paths from the root node to all leaf nodes in each independently trained decision tree. During the traversal, the features of each internal node along the path and its corresponding partition boundary value are completely recorded. The system unifies these continuous or discrete multi-dimensional threshold branch judgment logics into standardized "If-condition-Then-action" conditional expression rules, replacing the traditional complex rule system of a single tree. For CART classification trees corresponding to discrete control actions, the extracted expression form is such as "IF total load > 5000 AND electricity price = 0, THEN number of chillers in operation = 1"; while for CART regression trees corresponding to continuous control actions, the extracted expression form is such as "IF total load < 3000 AND outlet water temperature > 8℃, THEN pump frequency = 34Hz". Each device generates a simplified, non-redundant, and independent control rule subset corresponding to the HVAC physical mechanism.

[0049] Furthermore, after obtaining the independent conditional expression rules for each control action, a hierarchical summary and dynamic cascading rule integration is performed. Based on the standardized causal logic hierarchy framework 103 previously constructed based on Granger causality test 101 and physical constraints, the integration strictly follows the causal transmission sequence of the "cause → effect" driving relationship. Starting from the top-level driving device, the cascading rules are summarized and arranged level by level to the downstream response devices. During the dynamic activation of the entire site's rule cascading, the execution results of the conditional expressions generated in the preceding level (i.e., the output labels as effect variables) will be dynamically cascaded as the input for triggering and driving the conditional judgment of the conditional expressions in the subsequent response levels (downstream effect variables). The specific global cascade control logic and dynamic cascade relationship matching are as follows: The first-level magnetic levitation cooling quantity rule is driven only by the external total load and time-of-use electricity price. After its execution result is output, it serves as the dependent variable to trigger the cooling quantity rule of the centrifugal chiller belonging to the same first level, so as to achieve coordination within the same level of units and match the total cooling demand; subsequently, the execution result of the number of chillers started in the first level serves as the internal physical linkage characteristic of the preceding level, and is passed down and dynamically cascaded into the second-level dual-condition chillers and... In the heat pump cooling quantity rules, the trigger input for its conditional expression is determined by the load size and the first-level operating scale, which together determine the operating scale of the secondary peak-shaving heat pump. Then, the overall operating scale and operational status of each main cooling unit in the first and second levels are used as inputs to the cascaded third-level rules for the quantity of cold storage equipment and the opening / closing of the water tank valves. When the main cooling supply is sufficient, the conditional expression triggers the activation of the cold storage; when the main cooling supply is insufficient, the conditional expression triggers the release of cold storage. The system matches the causal logic of peak-valley cold storage and energy storage scheduling. The execution results of the cooling and energy storage equipment operation scale of the front-end host are jointly transmitted downwards and input into the outlet water temperature condition expression of the fourth level as the judgment benchmark for the heat exchange result state quantity. Finally, the execution results of the cooling scale and energy storage operation status of all hosts in the first, second, and third levels, together with the outlet water temperature execution results of the fourth level, are cascaded as the complete driving characteristics of the front-end level and input into the condition expressions of the water pump frequency and energy storage pump frequency in the fifth level at the bottom downstream. The system dynamically adjusts and increases the operating frequency of the water pump to enhance heat exchange according to the water temperature and the load of the host. Through this strategy integration method that strictly follows the causal logic topology architecture and hierarchical dynamic cascading, the coupling interference between control rules of different equipment and different attributes is completely cut off, eliminating the problems of incorrect control rule matching, deviation in operating condition prediction, or chaotic rule timing conflict. Finally, a set of global complete cascaded control rules for the energy station that takes into account physical causal consistency, high prediction accuracy, and clear and interpretable logic is constructed and output.

[0050] Example 5

[0051] Reference Figure 1 and Figure 2 This is the fifth embodiment of the present invention, which provides a method for extracting energy station operation strategy rules based on a multi-level decision tree with causal logic inference. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0052] A large regional energy station in Changzhou, Jiangsu Province, was selected as the test subject. This energy station has a total building area of ​​over 5.05 million square meters, encompassing office buildings, high-density residential buildings, and a large commercial complex. It is equipped with various heterogeneous equipment, including chillers, energy storage tanks, heat pumps, and water pumps. The test was conducted during the typical high-temperature cooling season in summer (July to August). In the pre-test preparation phase, an intelligent IoT sensing network deployed within the station was used to continuously collect data around the clock, with a uniform sampling step of one hour. This data included the station's total load, time-of-use electricity prices (peak, flat, and valley periods), real-time operating current of various units, start-up and shutdown status, water tank valve opening, and unit entry and exit data. A basic time-series operational dataset comprising 48 dimensions, including water temperature and pump operating frequency, was compiled, accumulating 1460 valid historical operational samples. In the detailed implementation process, the original time-series dataset was first corrected by dimensionless modeling and stationarity testing. Using the Granger causality test with physical unidirectional hierarchical constraints (101), all pairwise combinations of control variables were iterated. At this stage, the system strictly adhered to the unidirectional hierarchical constraint that "only higher-level variables are allowed to influence lower-level variables in HVAC physics." After filtering out spurious causal noise, a standardized causal logic hierarchy framework (103) containing a clear "drive-response" framework was constructed. The causal logic of the control actions is shown in Table 1.

[0053] Table 1. Causal Logic-Driven Correspondence Relationship of Control Actions

[0054] The multi-objective action decoupling and dedicated feature configuration module breaks down the complex station-wide linkage control task into multiple single-output control actions 201, such as discrete actions like the number of chillers and valve status, and continuous actions like pump frequency. It then configures a dedicated subset of strong causal feature inputs for each action according to a hierarchical framework. Subsequently, the classification and regression calculation array allocates independent decision trees 301 for single-objective training. Discrete actions use a CART classification tree with the Gini coefficient as the splitting criterion, while continuous actions use a CART regression tree with the minimum mean square error as the splitting criterion. During training, a pre-pruning strategy is introduced to limit the maximum tree depth and the minimum sample size of leaf nodes, thereby suppressing overfitting and simplifying the tree size. The global rule expression dynamic cascading integration module automatically traverses all paths from the root node to the leaf node of each decision tree, transforming them into standardized conditional expression rules. The execution results of the preceding driving layer are dynamically used as conditional judgment inputs for subsequent response layers. Finally, these rules are integrated into a complete global operating rule 401 that satisfies the HVAC physical coupling mechanism and is sent to the control bus, realizing closed-loop feedback linkage control of all station equipment.

[0055] To rigorously verify the superior performance of this invention compared to traditional technologies, under identical experimental boundary conditions and load requirements, a traditional control scheme without causal testing, which directly uses a single decision tree to fit the full dataset, was introduced as a control group. The accuracy of rule-condition matching between the two schemes in actual operation was compared and analyzed. The experimental data are shown in Table 2.

[0056] Table 2. Model Quantitative Evaluation Indicators: Rule Matching Accuracy

[0057] Analysis of the data in Table 2 shows that this invention has achieved breakthrough improvements in the rule fitting and operating condition matching accuracy of several core control actions. In the first-level basic cooling equipment control, the operating condition matching accuracy of the centrifugal chiller start-up and shutdown rule has increased from 88.83% to 92.49%. More significantly, in the key control actions involving the switching of cold and heat source supply / release modes and the scheduling of terminal hydraulic distribution, the accuracy of the water tank valve opening and closing rule has jumped significantly from 72.59% to 96.41%, achieving a precision leap of up to 23.82%. At the same time, the operating condition matching accuracy of the energy storage pump number rule, which is the core response quantity of the distribution system, has also been greatly improved from 65.30% to 95.62%, with a net improvement of 30.32%. The matching accuracy of the control rules directly determines the overall energy efficiency and operational stability of the system. This invention introduces a pre-emptive Granger causality test 101 for the first time in the feature engineering stage and creatively injects a unidirectional physical hierarchy constraint based on HVAC mechanisms. This mechanism, at both the mathematical statistics and engineering levels, completely eliminates the interference of spurious correlation features on the feature splitting path, enabling each decoupled single-objective control action to achieve high-precision focused fitting under a customized "pure" strong causal feature subset. By eliminating feature noise, the CART classification tree and regression tree can accurately capture the true physical boundary points of each device under critical transition conditions. This is the underlying scientific basis for the astonishing increases in accuracy of the control of the number of water tank valves and energy storage pumps, respectively, by 23.82% and 30.32%. Furthermore, through a dynamic cascaded integration module, the action results of the preceding driving level are dynamically used as input to the subsequent response level, constructing a vertically transmitted cascaded control mechanism on a global scale. This completely eliminates the branching interference and temporal action conflicts between continuous and discrete variable rules in traditional multi-objective synchronous fitting, ensuring that the global operation control rules possess unparalleled physical causal consistency and engineering robustness while maintaining extremely high operating condition prediction accuracy.

[0058] Example 6

[0059] Reference Figure 3 The sixth embodiment of the present invention provides an energy station operation strategy rule extraction system based on a causal logic inference multi-level decision tree, including a time series data preprocessing and causal topology inference module, a multi-objective action decoupling and exclusive feature configuration module, a classification regression calculation array and scale pre-pruning module, and a global rule expression dynamic cascading integration module.

[0060] The time-series data preprocessing and causal topology inference module is used to collect time-series operation data of energy station equipment, perform stabilization differential correction and causal verification, and inject unidirectional physical hierarchical constraints based on HVAC mechanism to construct a standardized causal logic hierarchical framework 103.

[0061] The multi-target action decoupling and exclusive feature configuration module is used to split the linkage control task into multiple independent single-output control actions 201 according to the physical action class, and based on the standardized causal logic hierarchy framework 103, it limits the selection to only external boundary features and driving device variables in the preceding level of the current control action, and constructs an exclusive feature set 202 for each single-output control action 201.

[0062] The classification and regression calculation array and the scale pre-pruning module are used to allocate independent decision trees 301 to independently train the exclusive feature set 202 of each single output control action 201. According to the action class, the classification splitting criterion is used for discrete actions and the regression splitting criterion is used for continuous actions. The pre-pruning strategy is introduced. By limiting the tree depth, constraining the sample size of leaf nodes and setting the minimum impurity reduction, the condition judgment level, eliminating working condition noise and merging low-discrimination branches are controlled respectively.

[0063] The global rule expression dynamic cascading integration module is used to traverse the decision paths of each decision tree in the computing array and uniformly transform them into standardized condition expression rules. Following the vertical transmission chain in the hierarchical framework, the execution results of the preceding driving hierarchical expression are dynamically used as the condition judgment input of the subsequent response hierarchical expression. Finally, it is integrated into a global complete operation rule 401 that satisfies the HVAC physical coupling mechanism and eliminates timing action conflicts.

[0064] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree proposed in the above embodiment.

[0065] This embodiment also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the energy station operation strategy rule extraction method based on causal logic inference multi-level decision tree proposed in the above embodiment.

[0066] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0067] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0068] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0069] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0070] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting energy station operation strategy rules based on a multi-level decision tree using causal logic inference, characterized in that, include: Granger causality test was performed on the time series data of control variables of each equipment in the energy station (101). Based on the test results, the unidirectional causal driving direction between each control variable was determined, and a Granger causal topology network was constructed (102). The network was then sorted topologically to determine the causal hierarchy and obtain a standardized causal logic hierarchy framework (103). The energy station control task is divided into multiple single-output control actions (201). Based on the standardized causal logic hierarchy framework (103), the operating status of the drive equipment located at the previous level of the current control action is selected as a feature to construct a unique feature set (202) for the single-output control action (201). For each single-output control action (201), an independent decision tree (301) is used for single-objective independent training, so that each independent decision tree (301) only fits the mapping relationship between the current control action and the corresponding feature set; Independent control rules corresponding to control actions are extracted from the independent decision tree (301), and the standardized causal logic hierarchical framework (103) is used as a guide to summarize and sort out the independent control rules in the order of causal transmission of each control action in the framework, thereby generating the complete operation rules of the entire energy station (401).

2. The method for extracting energy station operation strategy rules based on causal logic inference multi-level decision trees as described in claim 1, characterized in that: Before performing the Granger causality test (101), the time series data of the control variables of each device are tested for stationarity. For non-stationary time series data that fail the stationarity test, the time series trend is eliminated and the data is transformed into a standardized time series that meets the preset stationarity conditions, which serves as the data source for causal determination.

3. The method for extracting energy station operation strategy rules based on causal logic inference multi-level decision trees as described in claim 1 or 2, characterized in that: The construction of the Granger causal topology network (102) includes, During the causal verification process, physical unidirectional hierarchical constraints are applied, limiting the generation of causal driving edges from high-level control variables in the HVAC physics mechanism to low-level control variables, or allowing mutual influence between control variables at the same level. The physical unidirectional hierarchical constraints are used to correct the topology network generated by pure data-driven processes.

4. The method for extracting energy station operation strategy rules based on causal logic inference multi-level decision trees as described in claim 3, characterized in that: When constructing the exclusive feature set (202) of the single output control action (201), the operating status of the driving device located at the level preceding the current single output control action (201) in the standardized causal logic hierarchy framework (103) is selected as the internal physical linkage feature; System load variables and time-of-use electricity price variables, determined by the external environment and market mechanisms, are introduced as external boundary driving characteristics; The internal physical linkage features are combined with the external boundary driving features to construct a unique feature set for each single output control action (201).

5. The method for extracting energy station operation strategy rules based on causal logic inference multi-level decision trees as described in any one of claims 1, 2, and 4, characterized in that: When the independent decision tree (301) performs single-objective independent training, the single-output control action (201) that needs to be optimized is divided into discrete control action and continuous control action; For the discrete control action, a classification splitting criterion is used to train the corresponding independent decision tree (301) and output the control rules for the discrete category; For the continuous control action, the corresponding independent decision tree (301) is trained using the regression splitting criterion, and the continuous optimal setting value rule is output.

6. The method for extracting energy station operation strategy rules based on causal logic inference multi-level decision trees as described in claim 5, characterized in that: The independent decision tree (301) also includes the following when performing single-objective independent training: The independent training and growth process of the independent decision tree (301) is constrained by a pre-pruning constraint strategy, which includes at least one of limiting the maximum depth of each decision tree, constraining the minimum sample capacity of the leaf nodes, and setting a minimum impurity reduction amount. The maximum depth of each decision tree is limited to restrict the maximum number of linkage levels of conditional judgment sentences in the extracted independent control rules; The minimum sample capacity of the constrained leaf nodes is used to eliminate abnormal operating conditions in historical data with a probability lower than a preset threshold. The minimum impurity reduction is set so that tree growth is terminated early when the distinguishability of the control action under the precondition is lower than the preset gain, and the control state branches with low contribution are merged.

7. The method for extracting energy station operation strategy rules based on causal logic inference multi-level decision trees as described in claim 6, characterized in that: The generated energy station's complete operational rules (401) include, Traverse all decision paths from the root node to all leaf nodes in each independent decision tree (301), and transform the multi-dimensional threshold branch judgment logic in each decision path into standardized conditional expression rules; When summarizing and organizing at different levels, the execution results of the conditional expressions generated in the previous level are dynamically cascaded as the input for conditional judgment in the subsequent response levels, thereby achieving global cascaded control.

8. A system for extracting energy station operation strategy rules based on a causal logic inference multi-level decision tree, employing the energy station operation strategy rule extraction method based on a causal logic inference multi-level decision tree as described in any one of claims 1 to 7, characterized in that: It includes a time series data preprocessing and causal topology inference module (100), a multi-objective action decoupling and specific feature configuration module (200), a classification and regression calculation array and scale pre-pruning module (300), and a global rule expression dynamic cascade integration module (400). The time series data preprocessing and causal topology inference module (100) is used to collect time series operation data of energy station equipment, perform stabilization differential correction and causal verification, and inject unidirectional physical hierarchical constraints based on HVAC mechanism to construct a standardized causal logic hierarchical framework (103). The multi-target action decoupling and exclusive feature configuration module (200) is used to split the linkage control task into multiple independent single-output control actions (201) according to the physical action class, and according to the standardized causal logic hierarchy framework (103), it limits the selection to only external boundary features and driving device variables in the current control action's preceding level, and constructs an exclusive feature set (202) for each single-output control action (201). The classification and regression calculation array and the scale pre-pruning module (300) are used to allocate independent decision trees (301) to independently train the exclusive feature set (202) of each single output control action (201), and according to the action class division, the classification splitting criterion is adopted for discrete actions and the regression splitting criterion is adopted for continuous actions. The pre-pruning strategy is introduced, and by limiting the tree depth, constraining the sample size of leaf nodes and setting the minimum impurity reduction, the condition judgment level, eliminating working condition noise and merging low discrimination branches are controlled respectively. The global rule expression dynamic cascading integration module (400) is used to traverse the decision paths of each decision tree in the computing array and uniformly transform them into standardized condition expression rules. Following the vertical transmission chain in the hierarchical framework, the execution result of the preceding driving hierarchical expression is dynamically used as the condition judgment input of the subsequent response hierarchical expression, and finally integrated into a global complete operation rule (401) that satisfies the HVAC physical coupling mechanism and eliminates timing action conflicts.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for extracting energy station operation strategy rules based on causal logic inference multi-level decision tree as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for extracting energy station operation strategy rules based on causal logic inference multi-level decision tree as described in any one of claims 1 to 7.

Images