Campus energy consumption intelligent analysis system based on big data

Abstract

This invention discloses a campus energy consumption intelligent analysis system based on big data, belonging to the fields of big data analysis and energy management technology. It acquires multi-source energy consumption data from the campus and constructs a dynamic event graph. The system includes: a matrix construction module that builds a Bayesian causal network based on the dynamic event graph and historical energy consumption data, outputting a causal influence factor matrix; a candidate instruction module that generates a candidate instruction set based on the causal influence factor matrix using a genetic algorithm; a priority instruction module that outputs a priority instruction sequence based on the simulated energy consumption curve after executing the candidate instruction set; a collaborative optimization module that distributes the priority instruction sequence to each building node, generating device-level control instructions; and an autonomous strategy module that performs deep reinforcement learning optimization iterations based on the actual energy saving rate and comfort deviation feedback after executing the device-level control instructions. This invention effectively solves the inefficiency problems of IoT systems due to resource heterogeneity, rigid scheduling, and lack of autonomous optimization capabilities.

Description

Technical Field

[0001] This invention relates to the field of big data analytics and energy management technology, and in particular to a campus energy consumption intelligent analysis system based on big data. Background Technology

[0002] Currently, campus energy management generally faces systemic bottlenecks such as data fragmentation, extensive control, delayed response, and rigid strategies. These bottlenecks manifest themselves primarily at the following levels: First, at the data integration and analysis level, existing campus energy systems largely rely on various independently deployed smart meters and sensors. The types of data collected are limited, and spatiotemporal alignment is difficult, making it hard to form a panoramic view of the energy consumption-environment-event relationships. These systems often only achieve simple statistics and visualization of energy consumption data, lacking the ability to deeply integrate multi-source heterogeneous data and mine dynamic events, resulting in the data value not being fully realized.

[0003] Secondly, at the energy consumption control strategy level, existing management methods are mostly "open-loop" control based on fixed schedules or simple threshold triggers, such as timed switching of lighting and fixed temperature settings for air conditioning. These strategies cannot respond in real time to dynamic events such as personnel activities and environmental changes, nor can they quantify the actual causal impact of different events on energy consumption, resulting in both energy waste and decreased comfort. Thirdly, at the system coordination and optimization level, the campus lacks cross-regional and cross-system collaborative scheduling mechanisms. Even if some systems have local optimization capabilities, it is difficult to achieve optimal allocation of overall campus energy consumption. Finally, at the strategy adaptive evolution level, existing system control strategies rely heavily on manual experience settings, making it difficult to adapt to the challenges brought by complex factors such as equipment aging, changing usage habits, and abnormal weather during long-term operation. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a campus energy consumption intelligent analysis system based on big data. By constructing a full-link technology system encompassing "data perception - event mapping - causal inference - intelligent optimization - collaborative execution - strategy evolution," this invention achieves multi-dimensional deep perception of campus energy consumption, precise causal tracing, dynamic collaborative optimization, and continuous adaptive regulation. This effectively solves the problems of existing systems, such as insufficient situational awareness due to data fragmentation, coarse control due to unclear causal relationships, low overall energy efficiency due to lack of collaboration, and poor long-term adaptability due to rigid strategies.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution: a campus energy consumption intelligent analysis system based on big data, comprising the following steps:

[0008] The event analysis module is used to acquire multi-source energy consumption data on campus and construct a dynamic event graph of the campus.

[0009] The matrix construction module is used to build a Bayesian causal network based on the campus dynamic event map and historical energy consumption data, and output the causal influence factor matrix.

[0010] The candidate instruction module is used to generate a candidate instruction set based on the causal influence factor matrix using a genetic algorithm.

[0011] The priority instruction module is used to output a priority instruction sequence based on the pseudo-energy consumption curve after the candidate instruction set is executed.

[0012] The collaborative optimization module is used to allocate priority instruction sequences to each building node, negotiate quotas and synchronize status through the pre-built dynamic event graph of each building, and generate device-level control instructions.

[0013] The autonomous strategy module is used to perform deep reinforcement learning optimization and iteration based on the actual energy saving rate and comfort deviation feedback after the execution of device-level control commands, so as to obtain an adaptive control strategy.

[0014] Furthermore, acquire multi-source energy consumption data from the campus and construct a dynamic event graph of the campus, including:

[0015] Real-time readings of smart meters, smart water meters, and smart gas meters deployed at various metering points within campus buildings are collected, and environmental sensor data streams generated by temperature sensors, humidity sensors, light sensors, and personnel counting sensors deployed in the same physical space are collected simultaneously and aggregated into raw multi-source time-series data.

[0016] The original multi-source time series data is connected to a real-time stream processing platform, where missing value imputation is performed based on linear interpolation, and anomaly detection and correction are performed using a sliding window based on the three-times-standard-deviation principle, resulting in the output of cleaned standard time series data.

[0017] Based on a predefined event rule base and a sliding time window, real-time pattern matching and threshold judgment are performed on the standard time series data to extract energy consumption mutation events, environmental parameter over-limit events and personnel gathering events, and output an event feature set.

[0018] Using buildings, functional rooms, and energy-consuming equipment within the campus as entity nodes, and each event in the event feature set as event nodes, time-series edges are created according to the time sequence, and ownership edges are created according to spatial ownership. These are then instantiated and connected in the graph database to generate a basic campus event graph.

[0019] By inserting new nodes, deleting expired nodes, and updating edges and attributes on the basic campus event graph, a dynamic campus event graph is obtained.

[0020] Furthermore, based on a predefined event rule base and a sliding time window, real-time pattern matching and threshold judgment are performed on the standard time-series data to extract energy consumption mutation events, environmental parameter exceeding limit events, and personnel gathering events, outputting an event feature set, including:

[0021] Configure the length of the sliding time window and the sliding step size, and define the corresponding data source, feature calculation method, judgment threshold and minimum duration condition for each type of event in the event rule base;

[0022] Within each sliding window, the mean, variance, or gradient of the standard time series data covered by the window is calculated according to the event rules to obtain the current feature value sequence of each rule.

[0023] The current feature value sequence of each rule is compared with its predefined threshold, and the continuous time intervals in which the feature values ​​exceed the threshold are recorded as potential event intervals.

[0024] Verify whether the duration of each potential event interval meets the minimum duration condition of the corresponding rule, eliminate intervals that do not meet the condition, and determine the intervals that meet the condition as valid events;

[0025] For each valid event, associate its event type, time interval, spatial location, and feature intensity, encapsulate them in a data structure, and output the event feature set.

[0026] Furthermore, a Bayesian causal network is constructed based on the dynamic event graph and historical energy consumption data, outputting a causal influencing factor matrix, including:

[0027] Extract the type, timestamp, and intensity attributes of all event nodes from the campus dynamic event map, and obtain the building total energy consumption data sequence aligned with the timestamps of each event from the energy consumption history database. Merge the event attribute data and energy consumption data according to the time series to generate a multi-dimensional joint observation dataset containing event features and energy consumption labels.

[0028] Based on the multidimensional joint observation dataset, the network structure space is searched using the Bayesian information criterion scoring function and the hill-climbing search algorithm to obtain a directed acyclic graph representing the conditional dependency between event nodes and energy consumption nodes, which serves as the Bayesian network structure.

[0029] Based on the multidimensional joint observation dataset and the Bayesian network structure, the maximum likelihood estimation method is used to calculate the conditional probability table of each node in the network under its parent node combination, thus completing the parameter learning of the Bayesian causal network.

[0030] For each event node in the Bayesian causal network, do-calculus is applied to perform intervention. The conditional probability expectation of the target energy consumption node after intervention is calculated by subtracting the baseline expectation before intervention, and the causal effect value of the event node on energy consumption is obtained.

[0031] Summarize all event nodes and their corresponding causal effect values, construct a matrix with event nodes as rows and energy consumption status as columns, fill the causal effect values ​​into the corresponding positions, and output as a causal influence factor matrix.

[0032] Furthermore, the step of calculating the conditional probability table of each node in the network under its parent node combination using the maximum likelihood estimation method includes:

[0033] Based on the Bayesian network structure, all combinations of parent nodes for each node are determined to form multiple conditional configurations.

[0034] In the multidimensional joint observation dataset, the number of data samples falling under each condition configuration is counted.

[0035] For each node, under each conditional configuration, calculate the proportion of the number of samples of that node taking different state values ​​to the total number of samples under that conditional configuration;

[0036] All the calculated proportions are filled into the conditional probability table according to the correspondence between nodes and their parent node combinations.

[0037] Furthermore, a candidate instruction set is generated using a genetic algorithm based on the causal influence factor matrix, including:

[0038] Define the encoding rules for individual instructions, and randomly generate multiple individual instructions based on the encoding rules to form an initial instruction population;

[0039] Based on the causal effect values ​​in the causal influence factor matrix, the expected energy-saving effect and expected comfort impact of each instruction individual in the initial instruction population are calculated and combined into the fitness score of each instruction individual.

[0040] Based on the fitness score, instruction individuals with higher scores are selected from the initial instruction population to form the parent instruction population;

[0041] Perform a single-point crossover operation on the individual instructions in the parent instruction population, and perform a basic bit mutation operation on the crossover individual to generate a child instruction individual;

[0042] Merge the parent instruction population with the child instruction individuals to update it into a new generation instruction population;

[0043] Repeat the fitness score calculation, instruction individual screening, crossover mutation and population update operations until the set number of iterations is reached. The resulting new generation of instruction population is then decoded into specific device control instructions and output as the candidate instruction set.

[0044] Furthermore, based on the pseudo-energy consumption curves after the candidate instruction set is executed, a priority instruction sequence is output, including:

[0045] Each instruction in the candidate instruction set is decoded into a device control parameter, and the control parameter is input into a prediction model trained based on historical energy consumption data to obtain a future energy consumption prediction sequence for each instruction. The pseudo-energy consumption curve of the instruction is constructed based on the future energy consumption prediction sequence.

[0046] The predicted energy savings are calculated based on the simulated energy consumption curve and the current energy consumption benchmark curve, and the predicted comfort deviation is calculated based on the control parameters and the comfort benchmark parameters.

[0047] The predicted energy savings and predicted comfort deviations are normalized and then weighted and combined based on the energy saving priority and comfort priority of the system configuration to obtain the comprehensive evaluation score of each instruction.

[0048] Based on the comprehensive evaluation score, all individual instructions are sorted in descending order to generate an initial instruction sorting list;

[0049] The system detects execution time conflicts and device resource conflicts among individual instructions in the initial instruction sorting list. For instruction individuals with conflicts, priority decisions and execution time offset adjustments are made based on their comprehensive evaluation scores.

[0050] The list of individual instructions after conflict resolution and time adjustment is sorted according to the adjusted execution time order, and the output is a priority instruction sequence.

[0051] Furthermore, priority command sequences are allocated to each building node, and quota negotiation and status synchronization are performed through the pre-built dynamic event graph of each building to generate device-level control commands, including:

[0052] Based on the physical building to which the device controlled by each instruction in the priority instruction sequence belongs, the instructions are categorized to obtain a set of instructions to be executed for each building.

[0053] Based on the local campus dynamic event graph of each building node, the expected events and energy consumption demand in the future scheduling cycle are analyzed. Combined with the received set of instructions to be executed, the expected energy consumption quota and adjustable capacity of this building are calculated and submitted.

[0054] The expected energy consumption quotas and adjustable capacity submitted by all building nodes are summarized. The system's total energy consumption index is used as a constraint, and the demand satisfaction and energy-saving contribution of each building are used as optimization objectives to solve the problem and obtain the approved energy consumption quota and time window allocated to each building.

[0055] The approved energy consumption quota and time window are distributed to the corresponding building nodes, and the operation plan of their internal equipment is revised based on the local event graph.

[0056] Based on the revised operation plan, each building node parses each instruction in the set of instructions to be executed into the operation parameters of the specific controlled device within the time window, and generates device-level control instructions.

[0057] Furthermore, based on the feedback between the actual energy saving rate and comfort deviation after the execution of device-level control commands, deep reinforcement learning is used for optimization and iteration to obtain an adaptive control strategy, including:

[0058] The system collects actual energy consumption readings and indoor environmental parameters after executing the device-level control commands within an evaluation period. It calculates the actual energy saving rate based on the actual energy consumption readings and the benchmark energy consumption, and calculates the comprehensive comfort deviation based on the indoor environmental parameters and the preset comfort range.

[0059] The deviation between the actual energy saving rate and the overall comfort level is normalized and then weighted and summed according to the preset weight coefficients to calculate the instant reward value used to evaluate the effect of this instruction execution.

[0060] The system state obtained from the analysis of the current campus dynamic event graph, together with the executed instructions and actions, are used as inputs to a pre-built deep reinforcement learning value evaluation network, which then outputs a predicted value for future cumulative rewards.

[0061] Calculate the temporal difference error between the instantaneous reward value and the predicted value, and adjust the internal parameters of the value assessment network based on the temporal difference error using the gradient backpropagation algorithm;

[0062] Based on the prediction results output by the updated value evaluation network, the parameters of the policy network used to generate instructions are optimized, and an adaptive control policy is generated.

[0063] (III) Beneficial Effects

[0064] The campus energy consumption intelligent analysis system based on big data provided by this invention has the following beneficial effects:

[0065] 1. The system achieves deep fusion of multi-source data and precise perception and causal tracing of dynamic events. By integrating multi-dimensional heterogeneous data from electricity, water, gas, environment, and personnel, the system constructs a dynamic event map of the campus that evolves over time, transforming discrete readings into highly semantic "event knowledge." Furthermore, it innovatively utilizes Bayesian causal networks and do-calculus to quantify the true causal effects of various events on energy consumption, generating a causal influence factor matrix. This not only breaks down data silos and achieves panoramic situational awareness but also elevates the depth of analysis from correlation to causality, providing a scientific and quantitative basis for precise energy-saving regulation and control. It fundamentally overcomes the problem of crude control caused by data fragmentation and unclear causality in traditional systems.

[0066] 2. A closed loop of intelligent decision-making and refined execution, characterized by "global optimization and local collaboration," has been established. Based on causal analysis, the system uses a genetic algorithm to generate and evaluate candidate instruction sets with multiple objectives, forming a priority instruction sequence for global optimization. By distributing instructions to each building node and relying on the local event graph of each node for quota negotiation and status synchronization, finally, implementable device-level control instructions are generated. This mechanism achieves seamless integration from global energy efficiency target planning to local personalized adaptation, ensuring both the overall energy-saving effect of the campus and taking into account the real-time needs and operating status of each unit. It realizes cross-regional and cross-system collaborative optimal scheduling and refined control, solving the energy efficiency bottleneck caused by the lack of "open-loop" control and collaboration in the past.

[0067] 3. The system possesses autonomous policy evolution and long-term adaptability based on deep reinforcement learning. It introduces an autonomous policy evolution module, using the actual energy saving rate and comfort deviation after command execution as feedback to continuously optimize the control policy network through a deep reinforcement learning framework. This enables the system to learn autonomously from historical control experience, continuously adapting to equipment aging, changes in personnel behavior patterns, and external environmental disturbances, achieving online iteration and long-term optimization of the control policy. The system thus possesses the ability to leap from "rule-driven" to "learning evolution," ensuring high efficiency, robustness, and sustainability in long-term operation, and realizing a fundamental shift in campus energy management from static configuration to dynamic adaptation. Attached Figure Description

[0068] Figure 1 This is a functional block diagram of the campus energy consumption intelligent analysis system based on big data according to the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Please see Figure 1 This invention provides a campus energy consumption intelligent analysis system based on big data, comprising the following steps:

[0071] The event analysis module is used to acquire multi-source energy consumption data on campus and construct a dynamic event graph of the campus.

[0072] In this embodiment, acquiring multi-source energy consumption data on campus and constructing a dynamic event graph of the campus includes:

[0073] Real-time readings of smart meters, smart water meters, and smart gas meters deployed at various metering points within campus buildings are collected, and environmental sensor data streams generated by temperature sensors, humidity sensors, light sensors, and personnel counting sensors deployed in the same physical space are collected simultaneously and aggregated into raw multi-source time-series data.

[0074] The original multi-source time series data is connected to a real-time stream processing platform, where missing value imputation is performed based on linear interpolation, and anomaly detection and correction are performed using a sliding window based on the three-times-standard-deviation principle, resulting in the output of cleaned standard time series data.

[0075] Based on a predefined event rule base and a sliding time window, real-time pattern matching and threshold judgment are performed on the standard time series data to extract energy consumption mutation events, environmental parameter over-limit events and personnel gathering events, and output an event feature set.

[0076] Using buildings, functional rooms, and energy-consuming equipment within the campus as entity nodes, and each event in the event feature set as event nodes, time-series edges are created according to the time sequence, and ownership edges are created according to spatial ownership. These are then instantiated and connected in the graph database to generate a basic campus event graph.

[0077] By inserting new nodes, deleting expired nodes, and updating edges and attributes on the basic campus event graph, a dynamic campus event graph is obtained.

[0078] In detail, the first step is to synchronously collect multi-source data. Specifically, this involves collecting real-time energy consumption readings generated by smart meters, smart water meters, and smart gas meters at key metering points deployed in various buildings on campus, including but not limited to power distribution rooms, main water supply pipes, gas inlets, and circuits of various large energy-consuming equipment.

[0079] Simultaneously, environmental sensor data streams generated by temperature sensors, humidity sensors, light sensors, and personnel counting sensors deployed in the same or related buildings, floors, or functional rooms are collected. More specifically, all the above readings and data streams are aggregated according to a unified time base to form a raw multi-source time-series data set containing timestamps, device identifiers, spatial location identifiers, and various measurement values.

[0080] Preferably, the second step is to perform cleaning and standardization processing of the time-series data. Specifically, the raw multi-source time-series data is connected to the data receiving interface of a real-time stream processing platform. Specifically, in the platform's data processing pipeline, a missing value imputation operation based on linear interpolation is first performed. That is, for any breakpoints in the data stream, a linear estimate is calculated and inserted based on the values ​​and time interval of the two most recent valid data points before and after that data point to maintain the continuity of the data stream.

[0081] Next, anomaly detection and correction are performed using a sliding window based on the three-standard-deviation principle. The sliding window moves sequentially across the data stream at a preset fixed time length, calculating the arithmetic mean and standard deviation of the data within each window. Any data point deviating from the mean by more than three standard deviations is identified and marked as an outlier, and this outlier is replaced with the arithmetic mean of the remaining normal data points within the window or the value of the previous normal data point. More specifically, after completing the above cleaning operations, standard time-series data with a uniform format, continuity, and no obvious anomalies is output.

[0082] Preferably, the third step is to perform real-time event identification and feature extraction based on rules. Specifically, a predefined event rule library is invoked, which defines triggering conditions for energy consumption mutation events, environmental parameter exceeding limits events, and personnel gathering events, including the associated data source fields, feature calculation methods, judgment thresholds, and the minimum duration of the event.

[0083] Specifically, a sliding time window is set, the length of which and the sliding step size are configured according to the real-time requirements of event monitoring. Within each sliding window, the current feature value sequence corresponding to each rule is calculated for the standard time-series data covered by the window, according to the mean calculation method, variance calculation method, or gradient calculation method defined in the event rule base. The calculated current feature value sequence is compared one by one with the predefined judgment thresholds in the rule base, and the continuous time intervals in which the feature values ​​exceed the thresholds are recorded and marked as potential event intervals.

[0084] Then, the verification operation determines whether the duration of each potential event interval meets the minimum duration condition specified in the corresponding event rule. Intervals that do not meet the condition are eliminated, and intervals that meet all conditions are officially determined as valid events. More specifically, for each valid event, its event type, the start and end time interval of the event, the spatial location identifier of the building or room associated with the event, and the intensity value obtained from feature calculation are encapsulated. The final output is a structured set of event features containing multiple event records.

[0085] Preferably, the fourth step is to construct a basic event knowledge graph. Specifically, physical entities within the campus, including buildings, functional rooms, and major energy-consuming equipment, are created as entity nodes in the graph database, and each node is assigned a unique identifier and type attribute. Simultaneously, each valid event record in the event feature set is created as a corresponding type event node in the graph database, with node attributes including time, location, and intensity information.

[0086] Specifically, the operation of creating temporal edges involves traversing all event nodes and establishing edges with direction and temporal relationships between event nodes that have a temporal correlation or potential causal relationship, based on the chronological order of their timestamps. The operation of creating attribution edges involves linking event nodes to entity nodes representing that location (building, room, or equipment) based on the spatial location identifier recorded by the event node, establishing an attribution relationship from event to entity. More specifically, through the query and creation interface of the graph database, the instantiation and connection operations of the above nodes and edges are executed to generate a basic campus event graph stored in a graph structure, reflecting the relationships between entities and events within a specific time period.

[0087] Preferably, the fifth step is to implement dynamic maintenance and updating of the graph. Specifically, this process is a continuously running service. Specifically, it involves inserting new nodes: when a newly identified valid event occurs, following the procedures in steps three and four, a new event node and its associated edge with the entity node are created and linked to the relevant existing event node.

[0088] Execute expired node deletion operations: Set an event validity period, periodically scan the time attributes of all event nodes in the graph, and remove nodes whose occurrence time is earlier than the current time minus the validity period, along with their associated edges, from the graph database. Execute edge and attribute update operations: Based on the latest collected and processed data, refresh the attributes of existing nodes, such as updating the status description of entity nodes, or adding, deleting, or modifying temporal edges between nodes based on newly discovered event relationships.

[0089] More specifically, through the continuous insertion, deletion and update operations described above, the basic campus event map can dynamically evolve over time, ultimately resulting in and maintaining a dynamic campus event map that reflects campus energy consumption and environmental conditions in real time.

[0090] In this embodiment, based on a predefined event rule base and a sliding time window, real-time pattern matching and threshold judgment are performed on the standard time-series data to extract energy consumption mutation events, environmental parameter exceeding limit events, and personnel gathering events, and output an event feature set, including:

[0091] Configure the length of the sliding time window and the sliding step size, and define the corresponding data source, feature calculation method, judgment threshold and minimum duration condition for each type of event in the event rule base;

[0092] Within each sliding window, the mean, variance, or gradient of the standard time series data covered by the window is calculated according to the event rules to obtain the current feature value sequence of each rule.

[0093] The current feature value sequence of each rule is compared with its predefined threshold, and the continuous time intervals in which the feature values ​​exceed the threshold are recorded as potential event intervals.

[0094] Verify whether the duration of each potential event interval meets the minimum duration condition of the corresponding rule, eliminate intervals that do not meet the condition, and determine the intervals that meet the condition as valid events;

[0095] For each valid event, associate its event type, time interval, spatial location, and feature intensity, encapsulate them in a data structure, and output the event feature set.

[0096] Preferably, the first step is to configure the analysis parameters and event rules. Specifically, this step sets the operational framework and judgment criteria for subsequent real-time analysis. Specifically, it configures the length of the sliding time window and the sliding step size. The length determines the data time range examined each time data analysis is performed, for example, set to five minutes to capture short-term energy consumption fluctuations; the sliding step size determines the time interval for advancing the window, for example, set to one minute to achieve high-frequency scrolling detection.

[0097] Simultaneously, for each type of event to be detected—namely, energy consumption mutation events, environmental parameter exceeding limits events, and personnel gathering events—the corresponding data source, feature calculation method, judgment threshold, and minimum duration condition are defined in the event rule base. The data source specifies the specific sensor data fields upon which the analysis relies; the feature calculation method explicitly specifies whether mean calculation, variance calculation, or gradient calculation is used; the judgment threshold is a preset numerical critical point; and the minimum duration condition stipulates the shortest time for which the feature value must exceed the threshold before the event is confirmed. More specifically, this step establishes a quantifiable and adjustable detection rule foundation for monitoring abnormal states of different areas and types within the campus.

[0098] Preferably, the second step is to perform feature calculations within a sliding window. Specifically, this step transforms the raw data within the window into feature indicators that can be used for event determination. Specifically, within each sliding window, for the standard time-series data covered by the window, corresponding calculation operations are performed according to the feature calculation methods predefined for each type of event in the event rule base.

[0099] If the rule is defined as mean calculation, then the sum of all data points from the specified data source within the window is divided by the number of data points. If the rule is defined as variance calculation, then the average of the squares of the differences between each data point within the window and its mean is calculated. If the rule is defined as gradient calculation, then the ratio of the change in data values ​​at the beginning and end of the window to the time interval is calculated. More specifically, each calculation generates a current feature value characterizing the data characteristics of the window for the corresponding event rule. As the window continues to slide, the sequentially generated feature values ​​constitute the current feature value sequence of the rule.

[0100] Preferably, the third step is to identify potential event intervals based on thresholds. Specifically, this step uses a comparison operation to initially locate the time period of a suspicious event from a continuous sequence of feature values. Specifically, the current feature value sequence of each rule obtained in the second step is compared one by one with the predefined judgment threshold of that rule in the event rule base.

[0101] The alignment operation checks whether each feature value in the sequence is greater than or less than its corresponding threshold. The time intervals corresponding to all consecutive feature values ​​exceeding the threshold are recorded, and these consecutive time intervals are marked as potential event intervals. More specifically, this step completes the initial conversion from continuous numerical values ​​to discrete time intervals, filtering out candidate event intervals that require further verification.

[0102] Preferably, the fourth step is to verify and confirm valid events. Specifically, this step performs a persistence test on the initially selected intervals to eliminate transient interference and confirm genuine events. Specifically, for each potential event interval recorded in the third step, a verification operation is performed. The verification operation involves calculating the total duration of the potential event interval and comparing this duration with the minimum duration condition specified in the event rules corresponding to that interval.

[0103] If the duration of the interval is less than the minimum duration condition, a rejection operation is performed to remove the interval from the candidates. If the duration of the interval meets or exceeds the minimum duration condition, a confirmation operation is performed to officially confirm the interval as a valid event. More specifically, this step, through duration filtering, ensures the stability and significance of the extracted events, improving the accuracy of event detection.

[0104] Preferably, the fifth step is to encapsulate and output the event feature set. Specifically, this step assigns complete descriptive information to each confirmed event and forms a structured output. Specifically, for each valid event identified in the fourth step, its event type, time interval, spatial location, and feature intensity information are associated. The event type includes categories such as sudden energy consumption changes, exceeding environmental parameter limits, or personnel gatherings; the time interval is the start and end time of the event; the spatial location is the building and room identification where the sensor triggering the event is deployed; and the feature intensity can be the maximum amplitude or average value of the feature value exceeding a threshold within the event interval.

[0105] Subsequently, a data structure encapsulation operation is performed to integrate the aforementioned related information into a structured data record. More specifically, the data records of all currently valid events are aggregated, and the final output is a set of event features containing multiple event records, providing direct input for constructing a dynamic event graph.

[0106] The matrix construction module is used to build a Bayesian causal network based on the campus dynamic event map and historical energy consumption data, and output the causal influence factor matrix.

[0107] In this embodiment, a Bayesian causal network is constructed based on a dynamic event graph and historical energy consumption data, outputting a causal influence factor matrix, including:

[0108] Extract the type, timestamp, and intensity attributes of all event nodes from the campus dynamic event map, and obtain the building total energy consumption data sequence aligned with the timestamps of each event from the energy consumption history database. Merge the event attribute data and energy consumption data according to the time series to generate a multi-dimensional joint observation dataset containing event features and energy consumption labels.

[0109] Based on the multidimensional joint observation dataset, the network structure space is searched using the Bayesian information criterion scoring function and the hill-climbing search algorithm to obtain a directed acyclic graph representing the conditional dependency between event nodes and energy consumption nodes, which serves as the Bayesian network structure.

[0110] Based on the multidimensional joint observation dataset and the Bayesian network structure, the maximum likelihood estimation method is used to calculate the conditional probability table of each node in the network under its parent node combination, thus completing the parameter learning of the Bayesian causal network.

[0111] For each event node in the Bayesian causal network, do-calculus is applied to perform intervention. The conditional probability expectation of the target energy consumption node after intervention is calculated by subtracting the baseline expectation before intervention, and the causal effect value of the event node on energy consumption is obtained.

[0112] Summarize all event nodes and their corresponding causal effect values, construct a matrix with event nodes as rows and energy consumption status as columns, fill the causal effect values ​​into the corresponding positions, and output as a causal influence factor matrix.

[0113] Preferably, the first step involves extracting and merging data from the campus dynamic event map and historical database to construct a multi-dimensional joint observation dataset. Specifically, this step aims to align and fuse event information with energy consumption information along the time dimension, providing foundational data for subsequent causal analysis. Specifically, the type, timestamp, and intensity attributes of all event nodes are extracted from the campus dynamic event map. Event types include energy consumption mutation events, environmental parameter exceedance events, and personnel gathering events. The timestamp refers to the specific point in time when the event was detected. The intensity attribute is a numerical value that quantifies the severity or scale of the event.

[0114] Simultaneously, a sequence of total building energy consumption data aligned with the timestamps of each event is retrieved from the historical energy consumption database. This sequence records the total energy consumption values ​​of each building or area on campus at the same time. A merging operation is performed to align and merge the event attribute data and energy consumption data according to the time series, generating a multi-dimensional joint observation dataset containing event features and energy consumption labels. More specifically, each record in this dataset corresponds to a time point, containing the characteristics of various events occurring at that time and the energy consumption value at that time, thus establishing a temporal correlation between events and energy consumption.

[0115] Preferably, the second step involves searching for the optimal Bayesian network structure based on a multidimensional joint observation dataset. Specifically, this step aims to automatically learn the conditional dependencies between events and energy consumption variables from the data and represent them in the form of a directed graph. Specifically, based on the multidimensional joint observation dataset, a network structure space search is performed using the Bayesian information criterion scoring function and a hill-climbing search algorithm. The Bayesian information criterion scoring function is used to evaluate the balance between the fit of each candidate network structure to the data and the model complexity. The hill-climbing search algorithm is an algorithm that performs local optimization in a structure space consisting of all possible directed acyclic graphs.

[0116] During the search operation, the algorithm starts with an initial network structure and iteratively generates neighboring structures by adding, deleting, or reversing edges. Each neighboring structure is evaluated using a scoring function, and the structure with the highest score is selected as the current structure for the next step, until no higher-scoring neighboring structure can be found. More specifically, the final directed acyclic graph (DAG) with the highest score obtained from the search operation is a Bayesian network structure representing the conditional dependencies between event nodes and energy consumption nodes. This structure reveals which event variables may directly or indirectly affect energy consumption variables.

[0117] Preferably, the third step is to learn the parameters, i.e., the conditional probability table, of the Bayesian network based on the dataset and network structure. Specifically, this step aims to quantify the strength of dependencies between variables in the network structure, i.e., to calculate the probability that child nodes take different values ​​given a parent node state. Specifically, based on the multidimensional joint observation dataset and the Bayesian network structure, the maximum likelihood estimation method is used to calculate the conditional probability table of each node in the network under its parent node combination.

[0118] During computation, the first step is to determine all parent node combinations for each node based on the network structure. Then, the number of data samples falling under each parent node combination condition and the number of samples where child nodes take different state values ​​under that condition are counted in the dataset. Finally, the conditional probability is estimated by calculating the proportions. More specifically, after calculating the conditional probability table for all nodes, the parameter learning of the Bayesian causal network is complete, enabling the network to perform probabilistic inference.

[0119] Preferably, the fourth step is to apply intervention analysis to each event node in the Bayesian causal network to calculate the causal effect value. Specifically, this step aims to go beyond correlation analysis and quantitatively assess the causal impact of a specific event on campus energy consumption. Specifically, for each event node in the Bayesian causal network, do-calculus is applied to execute the intervention. Mathematically, the intervention is represented by forcibly setting an event node to a specific state while cutting off all edges pointing to that node, to simulate a scenario where the event state is forcibly changed externally.

[0120] After intervention, based on the learned conditional probability table, the expected value of the conditional probability of the target energy consumption node under the intervention conditions is calculated. The difference between this post-intervention expected value and the baseline expected value under the pre-intervention conditions (i.e., natural conditions) is used to calculate the causal effect value of the event node on energy consumption. More specifically, the causal effect value is a signed numerical value whose magnitude and direction indicate the amount by which a forced change in the event state will lead to an average increase or decrease in energy consumption.

[0121] Preferably, the fifth step is to summarize all causal effect values ​​to construct and output a causal influence factor matrix. Specifically, this step aims to organize the dispersed causal effect measurements into a structured matrix form that is easy for subsequent optimization modules to use. Specifically, it summarizes all event nodes and their corresponding causal effect values, constructing a matrix framework with event nodes as rows and energy consumption states as columns.

[0122] In this matrix, rows correspond to each analyzed event variable, and columns correspond to different states or energy consumption indicators of the energy consumption variable. The input operation involves filling the cell in the matrix where the row of the event node intersects with the column of the corresponding energy consumption state, based on the causal effect value of each event node calculated in step four. More specifically, the complete matrix obtained after completing all input operations is the causal influence factor matrix. This matrix clearly and quantitatively characterizes the impact of various observable events on energy consumption on campus from a causal perspective, providing a crucial basis for intelligent generation of control commands.

[0123] In this embodiment, the step of calculating the conditional probability table of each node in the network under its parent node combination using the maximum likelihood estimation method includes:

[0124] Based on the Bayesian network structure, all combinations of parent nodes for each node are determined to form multiple conditional configurations.

[0125] In the multidimensional joint observation dataset, the number of data samples falling under each condition configuration is counted.

[0126] For each node, under each conditional configuration, calculate the proportion of the number of samples of that node taking different state values ​​to the total number of samples under that conditional configuration;

[0127] All the calculated proportions are filled into the conditional probability table according to the correspondence between nodes and their parent node combinations.

[0128] Preferably, the first step is to determine the parent node combinations and condition configurations based on the network structure. Specifically, this step aims to clarify the preconditions for probability calculation for each node based on the structural relationships of the Bayesian causal network. Specifically, the Bayesian network structure is a directed acyclic graph representing the conditional dependencies between event nodes and energy consumption nodes. Based on this structure, the operation of determining all parent node combinations for each node is as follows: for each node in the network, identify all parent nodes that directly point to that node, and list all possible combinations of the state values ​​of these parent nodes.

[0129] For example, for a node representing "total building energy consumption," its parent nodes might include nodes for "crowd gathering events in a certain area" and "outdoor high temperature exceeding limits events." Therefore, the combinations of its parent nodes would include various state combinations such as "crowd gatherings occur and high temperature events occur," and "crowd gatherings occur but high temperature events do not occur." More specifically, each such combination of parent node states is defined as an independent condition configuration, thereby constructing a series of explicit probabilistic calculation scenarios for each node to be calculated in the network.

[0130] Preferably, the second step is to statistically analyze the sample frequency of each condition configuration in the multidimensional joint observation dataset. Specifically, this step utilizes historical observation data to quantify the empirical frequency of each parent node state combination. More specifically, the multidimensional joint observation dataset is formed by merging event attribute data from the campus dynamic event graph with energy consumption data from the historical energy consumption database, aligned and time-stamped.

[0131] The statistical operation involves: for each conditional configuration determined for each node in the previous step, traversing and matching throughout the entire multidimensional joint observation dataset, counting the number of data samples in which the actual value of the parent node variable exactly matches that conditional configuration. More specifically, this operation results in a specific number of data samples for each node and each conditional configuration in the network. This number directly reflects the frequency of co-occurrence of that particular parent node state in history, providing an empirical data basis for subsequent probability calculations.

[0132] Preferably, the third step is to calculate the empirical conditional probability of the node under different state values. Specifically, this step, based on the statistical results of the second step, calculates the probability proportion of the node taking each of its own states given a combination of parent node states.

[0133] Specifically, for each node in the network, a calculation operation is performed for each of its conditional configurations. First, within the subset of data samples corresponding to that conditional configuration, the number of samples in which the node takes each possible state value (e.g., the possible states of an energy-consuming node are "high," "medium," and "low") is counted. Next, the core of the calculation operation is to divide the number of samples in which the node takes a specific state value by the total number of samples under that conditional configuration, thus obtaining a ratio value between 0 and 1. More specifically, this ratio value represents the empirical conditional probability learned from historical data that the node is in a certain state under the constraints of a specific parent node's state.

[0134] Preferably, the fourth step is to construct and populate a complete conditional probability table. Specifically, this step systematically organizes the discrete probability values ​​calculated in the previous step into a structured probability distribution table. Specifically, a conditional probability table is created for each node in the Bayesian causal network.

[0135] The filling operation involves taking the proportions calculated in step three for each node under each conditional configuration and for each possible state value, and strictly following the correspondence between nodes, conditional configurations, and state values, and filling them one by one into the designated cells of the conditional probability table corresponding to that node. More specifically, after this filling operation, the conditional probability table of each node fully depicts how its state probability changes with the state of its parent node. All these conditional probability tables collectively complete the parameter learning of the Bayesian causal network, enabling it to be used for subsequent quantitative calculation of causal effects, thus providing core probabilistic model support for generating the causal influence factor matrix.

[0136] The candidate instruction module is used to generate a candidate instruction set based on the causal influence factor matrix using a genetic algorithm.

[0137] In this embodiment, a candidate instruction set is generated based on the causal influence factor matrix using a genetic algorithm, including:

[0138] Define the encoding rules for individual instructions, and randomly generate multiple individual instructions based on the encoding rules to form an initial instruction population;

[0139] Based on the causal effect values ​​in the causal influence factor matrix, the expected energy-saving effect and expected comfort impact of each instruction individual in the initial instruction population are calculated and combined into the fitness score of each instruction individual.

[0140] Based on the fitness score, instruction individuals with higher scores are selected from the initial instruction population to form the parent instruction population;

[0141] Perform a single-point crossover operation on the individual instructions in the parent instruction population, and perform a basic bit mutation operation on the crossover individual to generate a child instruction individual;

[0142] Merge the parent instruction population with the child instruction individuals to update it into a new generation instruction population;

[0143] Repeat the fitness score calculation, instruction individual screening, crossover mutation and population update operations until the set number of iterations is reached. The resulting new generation of instruction population is then decoded into specific device control instructions and output as the candidate instruction set.

[0144] Preferably, the first step is to define the encoding rules for individual instructions and construct an initial instruction population. Specifically, this step aims to establish a clear and computable method for representing individuals for the evolutionary process of the genetic algorithm and to generate an initial solution set. Specifically, the encoding rules for individual instructions are defined, the core of which is to map a specific device control instruction into a fixed-length binary encoded string. The encoded string is pre-divided into multiple consecutive fields, each corresponding to a control parameter.

[0145] For example, the first few bits of the encoding string are used to encode the unique identifier of the target building; subsequent bits are used to encode the type of equipment, such as air conditioning, lighting, or water pumps; the following bits are used to encode the specific control action, such as turning on, turning off, or adjusting; and the last bit is used to encode the parameter value of the action, such as the temperature setpoint or brightness percentage. More specifically, based on this encoding rule, a random generation operation is performed, that is, a random number generator is used to independently and randomly assign the value of zero or one to each bit of the encoding string, thereby generating a large number of individual instructions with random combinations in terms of control target, action, and parameters. Collecting these randomly generated individual instructions constitutes the initial instruction population for algorithm evolution, providing a broad initial exploration space for searching the optimal control strategy.

[0146] Preferably, the second step is to calculate the fitness score of each instruction individual based on the causal influence factor matrix. Specifically, this step aims to quantitatively evaluate the potential comprehensive value of the control scheme represented by each individual in the initial instruction population. Specifically, calculations are performed based on the causal effect values ​​in the causal influence factor matrix. This matrix quantifies the causal influence strength of different campus events on energy consumption status. The calculations first assess the expected energy-saving effect of each instruction individual, i.e., by parsing the encoding string of the instruction individual, decoding the equipment and actions to be intervened in, and mapping them to simulated interventions on specific event nodes; subsequently, combining the causal effect value of the event node on the total energy consumption node in the causal influence factor matrix, the potential energy consumption reduction after executing this instruction is inferred and predicted.

[0147] Simultaneously, the expected comfort impact of the instruction is calculated, that is, based on the decoded control parameters, the degree to which its execution might deviate from the preset comfort range in the indoor environmental parameters of the relevant area. More specifically, the calculated expected energy-saving effect value and the expected comfort impact value are combined according to a preset weighted merging rule, and finally, a single fitness score is assigned to each instruction individual. The level of this score directly represents the quality of the balance achieved by the individual in pursuing energy-saving potential and ensuring user comfort.

[0148] Preferably, the third step is to select parent instruction populations from the initial population based on fitness scores. Specifically, this step aims to perform the selection operation in the genetic algorithm, preserving potentially better solutions for reproduction. Specifically, the selection operation is performed based on the fitness scores calculated for each individual in the initial instruction population in the second step.

[0149] The selection process employs a roulette wheel selection algorithm, where each instruction individual is assigned a probability of selection proportional to its fitness score. Then, through random sampling, a fixed number of instruction individuals are selected from the initial instruction population according to this probability distribution. More specifically, this selected group of instruction individuals, representing the current optimal solution, is formally named the parent instruction population, and they will serve as the genetic basis for generating the next generation of new individuals.

[0150] Preferably, the fourth step involves performing crossover and mutation operations on the parent instruction population to generate offspring instruction individuals. Specifically, this step aims to explore new solution spaces by simulating gene recombination and mutation in biological heredity. Specifically, instruction individuals in the parent instruction population are randomly paired, and a single-point crossover operation is performed with a preset crossover probability. This operation involves randomly selecting a position within the length of the encoded string as the crossover point, and then exchanging all binary bit segments after the crossover point between the two paired parent individuals, thereby generating two new encoded strings that incorporate the genetic characteristics of both parents. Subsequently, basic bit mutation operations are performed on all newly generated and non-crossover individual encoded strings with a preset, low mutation probability.

[0151] This operation iterates through each bit of the encoded string and randomly flips the value of that bit with a small probability, either changing zero to one or one to zero. More specifically, through crossover and mutation operations, a batch of child instructions that inherit the excellent characteristics of their parent generation while introducing new variations are generated.

[0152] Preferably, the fifth step is to merge the populations and update them to generate a new generation of command populations. Specifically, this step aims to complete a generational shift in the population. Specifically, a merging operation is performed, combining the parent command populations selected in the third step with the newly generated offspring command individuals in the fourth step into a temporary mixed population. Next, an update operation is performed, based on a preset fixed population size, sorting individuals from this temporary mixed population according to their fitness scores from high to low, and selecting the top-ranked individuals whose numbers match the initial population size. More specifically, these selected superior individuals constitute the new generation of command populations, replacing the previous generation and preparing to enter the next evolutionary cycle.

[0153] Preferably, the sixth step is to iteratively evolve and decode to output the final candidate instruction set. Specifically, this step aims to drive the population to continuously evolve towards a better solution region through multiple rounds of selection, crossover, and mutation. Specifically, the new generation of instruction population is used as the new initial population, and the fitness score calculation operation in the second step, the instruction individual selection operation in the third step, the crossover and mutation operation in the fourth step, and the population update operation in the fifth step are repeatedly executed. This complete iterative cycle continues until the maximum number of iterations preset in the algorithm parameters is reached.

[0154] When the iteration terminates, for each individual instruction in the final generation of instruction population, the binary encoded string is decoded in reverse according to the encoding rules defined in the first step. The decoding operation involves querying a predefined mapping table to convert the binary values ​​of each field in the encoded string into specific building identifiers, equipment types, control actions, and parameter values, thereby restoring them into specific equipment control instructions that can be executed by the campus equipment control system.

[0155] More specifically, these decoded device control commands, representing the final optimization results of the algorithm, are gathered together and formally output as the candidate instruction set for subsequent priority ranking and collaborative optimization. The entire process achieves automated optimization from causal analysis to the generation of diverse, high-quality control strategies.

[0156] The priority instruction module is used to output a priority instruction sequence based on the pseudo-energy consumption curve after the candidate instruction set is executed.

[0157] In this embodiment, based on the pseudo-energy consumption curve after the candidate instruction set is executed, a priority instruction sequence is output, including:

[0158] Each instruction in the candidate instruction set is decoded into a device control parameter, and the control parameter is input into a prediction model trained based on historical energy consumption data to obtain a future energy consumption prediction sequence for each instruction. The pseudo-energy consumption curve of the instruction is constructed based on the future energy consumption prediction sequence.

[0159] The predicted energy savings are calculated based on the simulated energy consumption curve and the current energy consumption benchmark curve, and the predicted comfort deviation is calculated based on the control parameters and the comfort benchmark parameters.

[0160] The predicted energy savings and predicted comfort deviations are normalized and then weighted and combined based on the energy saving priority and comfort priority of the system configuration to obtain the comprehensive evaluation score of each instruction.

[0161] Based on the comprehensive evaluation score, all individual instructions are sorted in descending order to generate an initial instruction sorting list;

[0162] The system detects execution time conflicts and device resource conflicts among individual instructions in the initial instruction sorting list. For instruction individuals with conflicts, priority decisions and execution time offset adjustments are made based on their comprehensive evaluation scores.

[0163] Preferably, the first step is to decode the candidate instructions and generate a pseudo-energy consumption curve. Specifically, this step aims to transform the candidate instructions generated by the genetic algorithm into specific control parameters and evaluate their impact on future energy consumption after execution. Specifically, the binary encoded string of each individual instruction in the candidate instruction set is decoded according to a predefined encoding mapping relationship. The decoding operation involves querying a mapping table to sequentially convert the binary values ​​of each field in the encoded string into a specific target building identifier, controlled equipment type, specific control action, and detailed parameter values ​​of the action.

[0164] For example, an encoded string is decoded into an instruction to "adjust the air conditioning temperature in the laboratory on the fifth floor of the third teaching building to 26 degrees Celsius at 2 PM." These decoded device control parameters are then input into a pre-trained energy consumption prediction model. This model, trained based on historical campus energy consumption and environmental data, can simulate the building's energy consumption response under a given control strategy. After executing the input operation, the model outputs a sequence of predicted energy consumption values ​​for the corresponding instruction over a future scheduling cycle. More specifically, using this sequence as the vertical axis and future time points as the horizontal axis, a simulated energy consumption curve is plotted or constructed for the individual instruction. This curve visually displays the expected energy consumption change trajectory after executing the instruction.

[0165] Preferably, the second step is to calculate the predicted energy savings and the predicted comfort deviation. Specifically, this step aims to quantitatively assess the energy-saving potential and potential impact on comfort of each candidate instruction. Specifically, based on the simulated energy consumption curve and the current energy consumption baseline curve, calculations are performed to obtain the predicted energy savings. The current energy consumption baseline curve refers to the future energy consumption curve predicted based on the current operating mode and historical patterns without any special control intervention.

[0166] The calculation operation involves comparing the predicted energy consumption values ​​of the simulated energy consumption curve at each time point with the baseline energy consumption values ​​of the current energy consumption benchmark curve at the same time point, and summing up the energy consumption reductions at all time points throughout the entire evaluation period to obtain the total predicted energy savings. Simultaneously, based on the aforementioned control parameters and comfort benchmark parameters, a calculation operation is performed to obtain the predicted comfort deviation. The comfort benchmark parameters are preset ranges of indoor environmental parameters for different building functional areas, such as the temperature comfort range of an air-conditioned room in summer.

[0167] The calculation process involves comparing the control parameters in the instruction, such as the planned temperature value, with the corresponding comfort benchmark parameter range for the area, assessing the degree of deviation, and synthesizing the deviations of multiple environmental parameters, such as temperature and illuminance, to form a quantified predicted comfort deviation. More specifically, this step generates two key evaluation indicators for each instruction, characterizing its expected effects from the dimensions of energy saving and comfort, respectively.

[0168] Preferably, the third step is to normalize and weight the indicators to obtain a comprehensive evaluation score. Specifically, this step aims to integrate evaluation indicators with different dimensions and meanings into a unified evaluation standard. Specifically, the predicted energy savings and the predicted comfort deviation are normalized respectively. The normalization process uses methods such as linear transformation to map the values ​​of all candidate instructions for each indicator to the same numerical range, eliminating differences in the dimensions and absolute magnitude of the original values, making them comparable relative scores.

[0169] Next, based on the system's configured energy-saving and comfort priorities, a weighted merging operation is performed. These priorities are pre-set by the campus energy management department, reflecting different emphases on energy-saving targets and comfort guarantees under specific periods or policies. The weighted merging operation involves multiplying the normalized predicted energy savings score by the energy-saving priority weight, and multiplying the normalized predicted comfort deviation score by the comfort priority weight, then adding the two products. More specifically, the sum is the comprehensive evaluation score for each instruction, reflecting its overall performance under the current management strategy.

[0170] Preferably, the fourth step is to generate an initial instruction ranking list based on the comprehensive evaluation score. Specifically, this step aims to initially rank the candidate instructions based on the quantitative evaluation results. Specifically, based on the comprehensive evaluation score of each individual instruction calculated in the third step, a descending order sorting operation is performed. The descending order sorting operation arranges all candidate instructions according to their comprehensive evaluation scores from highest to lowest, forming an ordered list. More specifically, this list is called the initial instruction ranking list, where the top-ranked instructions represent the control schemes that perform best overall under the current evaluation system and are most worthy of priority consideration.

[0171] Preferably, the fifth step is to detect and adjudicate instruction conflicts to adjust the execution plan. Specifically, this step aims to ensure that instructions prioritized earlier in the order do not interfere with each other during actual execution and remain executable. Specifically, it detects execution time conflicts and device resource conflicts between individual instructions in the initial instruction order list. An execution time conflict refers to two or more instructions being scheduled to execute within the same time period. A device resource conflict refers to two or more instructions requiring the scheduling of the same physical device, such as the same air conditioning unit or lighting circuit.

[0172] For conflicting instructions, priority decisions and execution time offset adjustments are performed based on their comprehensive evaluation scores. Priority decision-making directly compares the comprehensive evaluation scores of conflicting instructions, retaining those with higher scores while those with lower scores may be marked for adjustment or canceled. Execution time offset adjustment, for those retained instructions with time conflicts, postpones or advances their planned execution time to the nearest conflict-free idle time window. More specifically, after conflict detection and adjustment, the execution time and resources of all instructions are coordinated.

[0173] Preferably, the sixth step is to generate the final priority instruction sequence. Specifically, this step aims to output a conflict-free, optimized list of instructions that can be immediately delivered for execution. Specifically, the list of individual instructions, after conflict resolution and time adjustment, is reordered according to their adjusted planned execution time order.

[0174] The sorting operation arranges all instructions into an ordered queue, following the chronological order from earliest to latest. More specifically, this ordered queue is the final output sequence of priority instructions, which considers both the overall performance of the instructions and the feasibility of actual execution, and can be directly sent to the collaborative optimization module for distributed deployment.

[0175] The collaborative optimization module is used to allocate priority instruction sequences to each building node, negotiate quotas and synchronize status through the pre-built dynamic event graph of each building, and generate device-level control instructions.

[0176] In this embodiment, a priority instruction sequence is allocated to each building node. Quota negotiation and state synchronization are performed through a pre-built dynamic event graph of each building to generate device-level control instructions, including:

[0177] Based on the physical building to which the device controlled by each instruction in the priority instruction sequence belongs, the instructions are categorized to obtain a set of instructions to be executed for each building.

[0178] Based on the local campus dynamic event graph of each building node, the expected events and energy consumption demand in the future scheduling cycle are analyzed. Combined with the received set of instructions to be executed, the expected energy consumption quota and adjustable capacity of this building are calculated and submitted.

[0179] The expected energy consumption quotas and adjustable capacity submitted by all building nodes are summarized. The system's total energy consumption index is used as a constraint, and the demand satisfaction and energy-saving contribution of each building are used as optimization objectives to solve the problem and obtain the approved energy consumption quota and time window allocated to each building.

[0180] The approved energy consumption quota and time window are distributed to the corresponding building nodes, and the operation plan of their internal equipment is revised based on the local event graph.

[0181] Based on the revised operation plan, each building node parses each instruction in the set of instructions to be executed into the operation parameters of the specific controlled device within the time window, and generates device-level control instructions.

[0182] Preferably, the first step is to categorize the instructions to each building node based on the physical affiliation of the equipment. Specifically, this step aims to initially organize and distribute the centrally generated optimization instructions according to spatial jurisdiction. Specifically, a categorization operation is performed based on the unique identifier of the physical building to which the equipment regulated by each instruction in the priority instruction sequence belongs.

[0183] The classification operation involves parsing the target building information embedded or associated in each instruction, such as the building code, and adding the instruction to a storage structure indexed by that building code. More specifically, after all instructions in the priority instruction sequence have been parsed and assigned, all instructions contained in each storage structure indexed by the building identifier constitute the set of instructions to be executed for that building. This step establishes clear task boundaries and data ownership for subsequent distributed, on-demand negotiation.

[0184] Preferably, the second step involves each building node calculating and submitting resource proposals based on its local event graph. Specifically, this step aims to drive each building node as an autonomous agent, participating in global resource planning based on its own situational awareness. Specifically, each building node calls and queries its locally pre-built and maintained campus dynamic event graph. Based on this graph, a parsing operation is performed, that is, analyzing recent event nodes associated with the building's entity node in the graph, inferring expected events within future scheduling cycles, such as large lectures or experimental courses, and estimating the resulting basic energy consumption requirements.

[0185] Subsequently, based on the received set of instructions to be executed, a calculation operation is performed. This calculation operation integrates local event predictions and external instruction requirements to quantify the expected energy consumption quota needed for the building to execute these instructions. Simultaneously, it assesses the additional adjustable capacity that the building can provide through strategy adjustments while meeting basic functional and comfort requirements. More specifically, each building node encapsulates the calculated expected energy consumption quota and adjustable capacity into a resource proposal and proactively submits it to the central collaborative optimization module. This step demonstrates the node's local awareness and decision-making capabilities.

[0186] Preferably, the third step is for the central module to aggregate proposals and perform global optimization. Specifically, this step aims to coordinate the needs of each node from a system-wide perspective, achieving optimal resource allocation under overall constraints. Specifically, the central collaborative optimization module summarizes the expected energy consumption quotas and adjustable capacities submitted by all building nodes, using a pre-set total energy consumption index for the campus-level system as a rigid constraint.

[0187] With the common optimization objective of maximizing the satisfaction of demand in each building and the overall energy-saving contribution, a solution operation is performed. This operation applies a multi-objective optimization algorithm to calculate a set of optimal energy consumption quotas allocated to each building and their corresponding execution time periods, while satisfying the total energy consumption constraint. More specifically, the output of the solution is the approved energy consumption quota allocated to each building and a specific time window for executing that quota. This step achieves a balance between global benefits and local needs.

[0188] Preferably, the fourth step is to distribute the global decision and drive local plan revisions. Specifically, this step aims to translate the central coordination decision into concrete, actionable solutions for each node. Specifically, the central collaborative optimization module distributes the approved energy consumption quotas and time windows for each building, obtained in the third step, to the corresponding building nodes via communication links.

[0189] After receiving the approved energy consumption quota and time window, each building node performs a revision operation based on the detailed internal status and event sequence reflected in the local dynamic event graph. The revision operation adjusts and refines the preliminary operating plans of each energy-consuming subsystem or major equipment within the building, ensuring that its final operating curve is constrained within the issued approved energy consumption quota and time window. More specifically, the revised plan results in a detailed operating plan that is bound to the global quota and deeply integrated with the local event situation.

[0190] Preferably, the fifth step is for each node to generate final device-level control instructions based on the revised plan. Specifically, this step aims to transform the operational plan, which integrates global quotas and local knowledge, into final control signals capable of driving physical device actions. Specifically, each building node performs final parsing and instantiation of each instruction in the set of instructions to be executed, based on its internally revised detailed operational plan.

[0191] The parsing operation maps the control targets in the instructions to unique address identifiers of specific controlled devices based on the precise time windows and equipment operating parameters determined in the plan. It then assigns specific operating parameter values ​​to each device within the specified time window, such as the set temperature of air conditioners, the on / off status of lighting circuits, and the operating frequency of water pumps. More specifically, after parsing and parameter binding, each instantiated instruction generates a device-level control instruction that can be directly issued to field controllers or actuators. The collection of all these device-level control instructions constitutes the final output of the collaborative optimization module. These instructions are sent to the campus IoT network for execution, thereby achieving refined and collaborative intelligent control of campus energy consumption.

[0192] The autonomous strategy module is used to perform deep reinforcement learning optimization and iteration based on the actual energy saving rate and comfort deviation feedback after the execution of device-level control commands, so as to obtain an adaptive control strategy.

[0193] In this embodiment, based on the feedback between the actual energy saving rate and comfort deviation after the execution of device-level control commands, deep reinforcement learning is used for optimization iteration to obtain an adaptive control strategy, including:

[0194] The system collects actual energy consumption readings and indoor environmental parameters after executing the device-level control commands within an evaluation period. It calculates the actual energy saving rate based on the actual energy consumption readings and the benchmark energy consumption, and calculates the comprehensive comfort deviation based on the indoor environmental parameters and the preset comfort range.

[0195] The deviation between the actual energy saving rate and the overall comfort level is normalized and then weighted and summed according to the preset weight coefficients to calculate the instant reward value used to evaluate the effect of this instruction execution.

[0196] The system state obtained from the analysis of the current campus dynamic event graph, together with the executed instructions and actions, are used as inputs to a pre-built deep reinforcement learning value evaluation network, which then outputs a predicted value for future cumulative rewards.

[0197] Calculate the temporal difference error between the instantaneous reward value and the predicted value, and adjust the internal parameters of the value assessment network based on the temporal difference error using the gradient backpropagation algorithm;

[0198] Based on the prediction results output by the updated value evaluation network, the parameters of the policy network used to generate instructions are optimized, and an adaptive control policy is generated.

[0199] Preferably, the first step is to collect feedback data and calculate the deviation between the actual energy saving rate and the overall comfort level. Specifically, this step aims to transform the execution effect of control commands in the physical world into quantifiable evaluation indicators. Specifically, after a preset evaluation period, such as a course cycle or 24 hours, a data collection operation is performed. This operation collects actual energy consumption readings from smart meters at each building's metering points, generated during the execution of the device-level control commands, and indoor environmental parameters, including temperature, humidity, and illuminance, from sensor terminals deployed in each indoor space, via the campus IoT network.

[0200] Next, the actual energy saving rate is calculated based on the actual energy consumption reading and a predefined baseline energy consumption. The baseline energy consumption is a reference value for comparison, predicted based on historical data from the same period and normal operating conditions. The calculation operation uses the percentage reduction in actual energy consumption relative to the baseline energy consumption as the actual energy saving rate to measure the energy-saving effect.

[0201] Simultaneously, the overall comfort deviation is calculated based on the collected indoor environmental parameters and the preset comfort parameter ranges for different functional rooms. The calculation assesses the degree to which the measured values ​​of each environmental parameter deviate from their preset comfort ranges, and integrates deviations from multiple dimensions such as temperature and humidity to form a comprehensive comfort deviation value that characterizes the overall impact on comfort. More specifically, this step transforms discrete sensor readings into two core quantitative indicators for evaluating the effectiveness of the control strategy.

[0202] Preferably, the second step is to normalize the performance indicators and calculate the immediate reward value using weighted averages. Specifically, this step aims to construct a unified reward signal that can guide the optimization direction of the reinforcement learning algorithm. Specifically, the actual energy-saving rate and the overall comfort deviation calculated in the first step are normalized respectively. The normalization operation uses a linear scaling method to transform these two indicators, which have different physical meanings and numerical ranges, into the same dimensionless numerical interval, such as between zero and one, making them comparable and additive.

[0203] Next, based on the weighting coefficients preset during system initialization that reflect management preferences, a weighted summation operation is performed. This operation involves multiplying the normalized actual energy-saving rate by a weighting coefficient representing the importance of the energy-saving target, and multiplying the normalized overall comfort deviation by a weighting coefficient representing the importance of comfort assurance. These two products are then added together. More specifically, the resulting scalar value is the immediate reward value used to evaluate the effectiveness of this instruction execution. This value directly and immediately reflects the effectiveness of the control action under the overall target.

[0204] Preferably, the third step involves inputting the system state and the actions performed into a value assessment network to obtain future return predictions. Specifically, this step aims to evaluate the long-term potential value of the current decision-making context and action combination. Specifically, the system state representation vector, obtained by parsing the current campus dynamic event map, is used as part of the input. This system state representation vector contains a digital abstraction of the characteristics, spatial distribution, and environmental background information of all active events at the current moment.

[0205] The action code of the device-level control command that triggered this feedback is used as another part of the input. The input operation is performed, and the combined input data, consisting of the system state representation vector and the action code, is fed into the input layer of a pre-built deep reinforcement learning value evaluation network. This value evaluation network is a deep neural network model. More specifically, through its internal multi-layer nonlinear transformation forward propagation process, the network ultimately produces a scalar value at the output layer. This scalar value is the network's prediction of the total cumulative reward that can be obtained in the future after taking the command action in the current system state.

[0206] Preferably, the fourth step is to calculate the temporal difference error and update the parameters of the value assessment network accordingly. Specifically, this step aims to correct the parameters of the value assessment network by utilizing the difference between real-time feedback and prediction, making its predictions closer to reality. Specifically, the difference between the immediate reward value obtained in the second step and the predicted value of future cumulative rewards output by the value assessment network in the third step is calculated; this difference is the temporal difference error.

[0207] Temporal difference error quantifies the inaccuracy of the network's current predictions. Based on this temporal difference error, the internal weight parameters of the value assessment network are adjusted using a gradient backpropagation algorithm. The adjustment operation first calculates the loss function based on the temporal difference error, then calculates the gradient of the loss function with respect to the adjustable parameters of each layer of the network through backpropagation, and finally updates the values ​​of these parameters along the gradient descent direction. More specifically, this parameter adjustment operation aims to minimize the error of future predictions, enabling the value assessment network to more accurately evaluate the long-term value of different state-action pairs.

[0208] Preferably, the fifth step is to optimize the policy network and generate an adaptive policy based on the updated value network. Specifically, this step aims to directly optimize the control policy generation mechanism using improved value evaluation criteria. Specifically, the internal parameters of the policy network used to generate instructions are optimized based on the prediction results provided by the updated and more predictive value evaluation network.

[0209] The policy network is another deep neural network whose function is to receive the system state representation vector as input and directly output the probability distribution of the suggested device control commands. The optimization operation uses the predictions from the value assessment network as a guiding signal, calculates the update gradient of the policy network parameters using the policy gradient algorithm, and adjusts its parameters so that the policy network is more likely to output commands that will obtain higher predictions from the value assessment network in the future.

[0210] More specifically, through continuous iterative cycles of data collection and feedback, reward calculation, value network updates, and policy network optimization, the policy network's decision-making capabilities continuously evolve, ultimately converging into an adaptive control policy capable of autonomously generating high-performance device-level control commands based on real-time campus dynamic event graphs. This policy is encapsulated into an executable model file, used to replace or assist in some rule-based decisions in subsequent scheduling cycles, enabling the system's self-learning and continuous performance improvement.

[0211] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0212] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0213] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A campus energy consumption intelligent analysis system based on big data, characterized in that: include: The event analysis module is used to acquire multi-source energy consumption data on campus and construct a dynamic event graph of the campus. The matrix construction module is used to build a Bayesian causal network based on the campus dynamic event map and historical energy consumption data, and output the causal influence factor matrix. The candidate instruction module is used to generate a candidate instruction set based on the causal influence factor matrix using a genetic algorithm. The priority instruction module is used to output a priority instruction sequence based on the pseudo-energy consumption curve after the candidate instruction set is executed. The collaborative optimization module is used to allocate priority instruction sequences to each building node, negotiate quotas and synchronize status through the pre-built dynamic event graph of each building, and generate device-level control instructions. The autonomous strategy module is used to perform deep reinforcement learning optimization and iteration based on the actual energy saving rate and comfort deviation feedback after the execution of device-level control commands, so as to obtain an adaptive control strategy.

2. The campus energy consumption intelligent analysis system based on big data according to claim 1, characterized in that, Acquire multi-source energy consumption data on campus and construct a dynamic event graph of the campus, including: Real-time readings of smart meters, smart water meters, and smart gas meters deployed at various metering points within campus buildings are collected, and environmental sensor data streams generated by temperature sensors, humidity sensors, light sensors, and personnel counting sensors deployed in the same physical space are collected simultaneously and aggregated into raw multi-source time-series data. The original multi-source time series data is connected to a real-time stream processing platform, where missing value imputation is performed based on linear interpolation, and anomaly detection and correction are performed using a sliding window based on the three-times-standard-deviation principle, resulting in the output of cleaned standard time series data. Based on a predefined event rule base and a sliding time window, real-time pattern matching and threshold judgment are performed on the standard time series data to extract energy consumption mutation events, environmental parameter over-limit events and personnel gathering events, and output an event feature set. Using buildings, functional rooms, and energy-consuming equipment within the campus as entity nodes, and each event in the event feature set as event nodes, time-series edges are created according to the time sequence, and ownership edges are created according to spatial ownership. These are then instantiated and connected in the graph database to generate a basic campus event graph. By inserting new nodes, deleting expired nodes, and updating edges and attributes on the basic campus event graph, a dynamic campus event graph is obtained.

3. The campus energy consumption intelligent analysis system based on big data according to claim 2, characterized in that, Based on a predefined event rule base and a sliding time window, the standard time-series data undergoes real-time pattern matching and threshold judgment to extract energy consumption mutation events, environmental parameter exceedance events, and personnel gathering events, outputting a set of event features, including: Configure the length of the sliding time window and the sliding step size, and define the corresponding data source, feature calculation method, judgment threshold and minimum duration condition for each type of event in the event rule base; Within each sliding window, the mean, variance, or gradient of the standard time series data covered by the window is calculated according to the event rules to obtain the current feature value sequence of each rule. The current feature value sequence of each rule is compared with its predefined threshold, and the continuous time intervals in which the feature values ​​exceed the threshold are recorded as potential event intervals. Verify whether the duration of each potential event interval meets the minimum duration condition of the corresponding rule, eliminate intervals that do not meet the condition, and determine the intervals that meet the condition as valid events; For each valid event, associate its event type, time interval, spatial location, and feature intensity, encapsulate them in a data structure, and output the event feature set.

4. The campus energy consumption intelligent analysis system based on big data according to claim 1, characterized in that, A Bayesian causal network is constructed based on dynamic event graphs and historical energy consumption data, outputting a causal influence factor matrix, including: Extract the type, timestamp, and intensity attributes of all event nodes from the campus dynamic event map, and obtain the building total energy consumption data sequence aligned with the timestamps of each event from the energy consumption history database. Merge the event attribute data and energy consumption data according to the time series to generate a multi-dimensional joint observation dataset containing event features and energy consumption labels. Based on the multidimensional joint observation dataset, the network structure space is searched using the Bayesian information criterion scoring function and the hill-climbing search algorithm to obtain a directed acyclic graph representing the conditional dependency between event nodes and energy consumption nodes, which serves as the Bayesian network structure. Based on the multidimensional joint observation dataset and the Bayesian network structure, the maximum likelihood estimation method is used to calculate the conditional probability table of each node in the network under its parent node combination, thus completing the parameter learning of the Bayesian causal network. For each event node in the Bayesian causal network, do-calculus is applied to perform intervention. The conditional probability expectation of the target energy consumption node after intervention is calculated by subtracting the baseline expectation before intervention, and the causal effect value of the event node on energy consumption is obtained. Summarize all event nodes and their corresponding causal effect values, construct a matrix with event nodes as rows and energy consumption status as columns, fill the causal effect values ​​into the corresponding positions, and output as a causal influence factor matrix.

5. The campus energy consumption intelligent analysis system based on big data according to claim 4, characterized in that, The method of calculating the conditional probability table of each node in the network under its parent node combination using the maximum likelihood estimation method includes: Based on the Bayesian network structure, all combinations of parent nodes for each node are determined to form multiple conditional configurations. In the multidimensional joint observation dataset, the number of data samples falling under each condition configuration is counted. For each node, under each conditional configuration, calculate the proportion of the number of samples of that node taking different state values ​​to the total number of samples under that conditional configuration; All the calculated proportions are filled into the conditional probability table according to the correspondence between nodes and their parent node combinations.

6. The campus energy consumption intelligent analysis system based on big data according to claim 1, characterized in that, Candidate instruction sets are generated using a genetic algorithm based on the causal influence factor matrix, including: Define the encoding rules for individual instructions, and randomly generate multiple individual instructions based on the encoding rules to form an initial instruction population; Based on the causal effect values ​​in the causal influence factor matrix, the expected energy-saving effect and expected comfort impact of each instruction individual in the initial instruction population are calculated and combined into the fitness score of each instruction individual. Based on the fitness score, instruction individuals with higher scores are selected from the initial instruction population to form the parent instruction population; Perform a single-point crossover operation on the individual instructions in the parent instruction population, and perform a basic bit mutation operation on the crossover individual to generate a child instruction individual; Merge the parent instruction population with the child instruction individuals to update it into a new generation instruction population; Repeat the fitness score calculation, instruction individual screening, crossover mutation and population update operations until the set number of iterations is reached. The resulting new generation of instruction population is then decoded into specific device control instructions and output as the candidate instruction set.

7. The campus energy consumption intelligent analysis system based on big data according to claim 1, characterized in that, Based on the pseudo-energy consumption curve after the candidate instruction set is executed, the preferred instruction sequence is output, including: Each instruction in the candidate instruction set is decoded into a device control parameter, and the control parameter is input into a prediction model trained based on historical energy consumption data to obtain a future energy consumption prediction sequence for each instruction. The pseudo-energy consumption curve of the instruction is constructed based on the future energy consumption prediction sequence. The predicted energy savings are calculated based on the simulated energy consumption curve and the current energy consumption benchmark curve, and the predicted comfort deviation is calculated based on the control parameters and the comfort benchmark parameters. The predicted energy savings and predicted comfort deviations are normalized and then weighted and combined based on the energy saving priority and comfort priority of the system configuration to obtain the comprehensive evaluation score of each instruction. Based on the comprehensive evaluation score, all individual instructions are sorted in descending order to generate an initial instruction sorting list; The system detects execution time conflicts and device resource conflicts among individual instructions in the initial instruction sorting list. For instruction individuals with conflicts, priority decisions and execution time offset adjustments are made based on their comprehensive evaluation scores. The list of individual instructions after conflict resolution and time adjustment is sorted according to the adjusted execution time order, and the output is a priority instruction sequence.

8. The campus energy consumption intelligent analysis system based on big data according to claim 1, characterized in that, Priority command sequences are allocated to each building node. Quota negotiation and status synchronization are performed using a pre-built dynamic event graph for each building, generating device-level control commands, including: Based on the physical building to which the device controlled by each instruction in the priority instruction sequence belongs, the instructions are categorized to obtain a set of instructions to be executed for each building. Based on the local campus dynamic event graph of each building node, the expected events and energy consumption demand in the future scheduling cycle are analyzed. Combined with the received set of instructions to be executed, the expected energy consumption quota and adjustable capacity of this building are calculated and submitted. The expected energy consumption quotas and adjustable capacity submitted by all building nodes are summarized. The system's total energy consumption index is used as a constraint, and the demand satisfaction and energy-saving contribution of each building are used as optimization objectives to solve the problem and obtain the approved energy consumption quota and time window allocated to each building. The approved energy consumption quota and time window are distributed to the corresponding building nodes, and the operation plan of their internal equipment is revised based on the local event graph. Based on the revised operation plan, each building node parses each instruction in the set of instructions to be executed into the operation parameters of the specific controlled device within the time window, and generates device-level control instructions.

9. The campus energy consumption intelligent analysis system based on big data according to claim 1, characterized in that, Based on the feedback between the actual energy saving rate and the comfort deviation after the execution of device-level control commands, deep reinforcement learning is used for optimization and iteration to obtain an adaptive control strategy, including: The actual energy consumption readings and indoor environmental parameters are collected after the device-level control commands are executed within an evaluation period. The actual energy saving rate is calculated based on the actual energy consumption readings and the benchmark energy consumption. The comprehensive comfort deviation is calculated based on the indoor environmental parameters and the preset comfort range. The deviation between the actual energy saving rate and the overall comfort level is normalized and then weighted and summed according to the preset weight coefficients to calculate the instant reward value used to evaluate the effect of this instruction execution. The system state obtained from the analysis of the current campus dynamic event graph, together with the executed instructions and actions, are used as inputs to a pre-built deep reinforcement learning value evaluation network, which then outputs a predicted value for future cumulative rewards. Calculate the temporal difference error between the instantaneous reward value and the predicted value, and adjust the internal parameters of the value assessment network based on the temporal difference error using the gradient backpropagation algorithm; Based on the prediction results output by the updated value evaluation network, the parameters of the policy network used to generate instructions are optimized, and an adaptive control policy is generated.

Images