Central air conditioning source and load collaborative control method and system

Abstract

The present application relates to a kind of central air conditioning source and load collaborative control method and system, for the data isomerism in the regulation and control process of existing chilled water system, semantic is difficult to identify, parameter self-adapting difficulty is high and energy efficiency constraint is difficult to dynamically satisfy, a kind of method based on multi-source data standardization, space-time alignment, semantic annotation and graph embedding is proposed, by constructing load semantic graph and introducing context-aware representation vector, the accurate extraction of multi-dimensional feature to end operating state and semantic reasoning are realized.Combining pre-training nonlinear mapping network with dynamic safety clipping, generate fuzzy PID control parameters that meet the system energy efficiency bandwidth constraints, and through hierarchical fuzzy rule base adaptive collaborative adjustment water supply temperature and pump group frequency.The method has adaptive closed-loop optimization capability, can improve system regulation accuracy, response speed and operating energy efficiency, effectively make up for the deficiency of traditional rule base and static parameter, promote equipment operation more stable, efficient and intelligent.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of central air conditioning systems, and in particular to a method and system for coordinated control of central air conditioning source and load. Background Technology

[0002] Central air conditioning source-load coordinated control systems have been widely researched and deployed in building intelligence, energy efficiency optimization, and comfort management. Currently, the mainstream technical solutions in the field of intelligent central air conditioning control are represented by multi-level adaptive control architectures, typically employing a combination of distributed end-point demand identification and centralized source-side energy allocation. Key technical methods include PID parameter tuning based on fuzzy control, model predictive control (MPC), expert rule bases, and empirical formula parameter tables. Fuzzy PID control is widely used due to its flexible adjustment capabilities for complex nonlinear systems, making it suitable for dynamic, model-free, time-varying, and strongly coupled operating conditions of air conditioning systems. Model predictive control emphasizes global optimization, explicit constraint handling, and rolling optimization capabilities, making it suitable for scenarios involving multi-variable energy consumption allocation and comfort constraints. In recent years, the industry has continuously promoted intelligent optimization based on multi-source data fusion, such as introducing big data behavioral modeling, spatial heat map analysis, and zonal energy control.

[0003] Existing central air conditioning adaptive control technologies mainly revolve around the following aspects: First, loop-level regulation based on fuzzy logic and intelligent PID parameter self-tuning. This involves inputting physical quantities such as temperature difference deviation signals (e) and deviation change rate (ec), supplemented by rules set based on human experience, to achieve offline parameter presetting and online fine-tuning. Second, the introduction of MPC optimized scheduling, which predicts load changes globally and dynamically distributes energy consumption indicators to source-side or terminal devices. However, in practical engineering, fuzzy PID and MPC are often independent or weakly coupled. Parameter tuning largely relies on manual adjustments, lacking semantic mining and contextual recognition of terminal load behavior, making it difficult to achieve efficient mapping between physical deviation signals and system behavioral intentions. Typical application scenarios include the alternating load switching in multi-functional buildings such as reception areas, office areas, and meeting areas. Traditional controllers struggle to capture the complex demand evolution across regions and time periods.

[0004] Existing control systems generally suffer from the following technical defects and shortcomings: I. Fuzzy control strategies and model predictive control (MPC) lack synergistic optimization. Fuzzy PID and MPC control modules operate independently and are based on closed-loop feedback of a single physical deviation signal. Parameter tuning is limited by empirical rules in two-dimensional (e-ec) space. There is a lack of in-depth understanding of the load intentions of multiple areas and behaviors of the building, which easily leads to control command conflicts or energy allocation resource waste between the two.

[0005] Second, the load identification dimension is too narrow. It fails to extract and incorporate the diverse end-point demand characteristics (such as rapid cooling, steady-state maintenance, and nighttime hibernation) from a behavioral semantic perspective into the parameter adaptive tuning process, thus failing to achieve context-adaptive matching of parameters in complex scenarios.

[0006] Third, manually constructing expert rule bases is difficult to cover a large number of operating condition changes. When the thermal capacity of building structures, personnel distribution, and frequency of use of functional areas change, traditional control parameters and rule bases are fixed, resulting in poor adaptability and high maintenance costs, which seriously affect the long-term comfort and energy efficiency of the overall system.

[0007] Fourth, if the adjustment of control parameters relies solely on past or current individual sensor physical quantities, it will not adequately detect the temporal and causal effects of changes in the regional environment, resulting in delayed response or even amplification of disturbances. This can lead to system overshoot, oscillation, or frequent start-stop, affecting equipment lifespan and safety.

[0008] Fifth, the traditional parameter mapping process is a black box, parameter correction lacks interpretability and transparency, and control strategies are difficult to dynamically optimize for different building scenarios and operational goals, which is not conducive to the intelligent upgrading of actual engineering operation and maintenance.

[0009] Given the limitations of existing technologies, the core issues that the central air conditioning source-load coordinated control system urgently needs to overcome include: how to abstract complex energy consumption behaviors at the terminal into computable semantic features, construct a load semantic map to achieve contextual linkage perception, and then replace traditional manual rule tuning with data-driven nonlinear parameter mapping, enabling the fuzzy PID parameter tuning process to have cross-level, scenario self-consistency, and system-level collaborative capabilities; and how to adjust control parameters in real time for dynamically changing operating conditions to achieve efficient, robust, and interpretable control response for source-load coordination. These are precisely the technical deficiencies that this patented technology aims to address, and are of great significance for promoting the in-depth development of intelligent building energy conservation and intelligent electromechanical systems. Summary of the Invention

[0010] This application provides a method and system for coordinated control of central air conditioning source and load, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0011] This application provides a central air conditioning source-load coordinated control method and system, specifically including: S1: Acquire the raw data stream collected by the multi-source sensors at the central air conditioning terminal. The raw data stream includes the indoor temperature setpoint and actual temperature difference, fan coil unit start / stop status, water valve opening degree and timestamp information. Perform spatiotemporal alignment and cleaning processing on the raw data stream to generate a standardized terminal energy consumption characteristic sequence.

[0012] S2: Based on the standardized end-use energy consumption feature sequence, combined with building functional zoning information and personnel activity heat map distribution, semantic annotation processing is performed on the time-series segments in the original data stream. Basic semantic units including rapid cooling demand, steady-state maintenance demand, local overcooling suppression and nighttime low-load hibernation are defined, and a basic semantic tag set with spatial weight attributes is generated.

[0013] S3: Utilizing the causal relationships between the basic semantic tag set and historical operational data, an initial load semantic graph is constructed. Graph nodes represent the basic semantic units, and graph edges represent semantic co-occurrence relationships, temporal transition probabilities, and causal influence strength. The constructed initial load semantic graph is a topological model reflecting the dynamic evolution path of building loads.

[0014] S4: Dynamically embed the standardized end-point energy consumption feature sequence updated in real time into the initial load semantic map, identify the dominant semantic nodes and their neighborhood activation patterns under the current operating conditions through a lightweight graph neural network model, and calculate and output a load semantic representation vector with context awareness.

[0015] S5: Based on the load semantic representation vector and the pre-trained nonlinear mapping network, perform dynamic mapping operation of fuzzy PID control parameters. According to the thermal inertia level and response urgency in the node attributes of the initial load semantic map, calculate the direction and amplitude range of the proportional coefficient increment, integral coefficient increment and derivative coefficient increment, and generate a fuzzy PID parameter candidate set to be pruned.

[0016] S6: Based on the energy efficiency bandwidth constraints of the current chilled water system, perform safe pruning on the candidate set of fuzzy PID parameters to be pruned, eliminate parameter combinations that exceed the physical execution limit, and generate the final fuzzy PID control parameter set that meets the system stability requirements.

[0017] S7: Input the final fuzzy PID control parameter set into the fuzzy inference engine to replace the traditional fixed rule base based on error and deviation change rate. Automatically generate hierarchical fuzzy rules according to the hierarchical attributes of the nodes in the initial load semantic graph, and execute the output operation of water supply temperature setpoint adjustment and pump group frequency adjustment command.

[0018] S8: Monitor the jump status of the dominant semantic node and the distribution deviation of the neighborhood activation mode within the continuous sampling period. If the distribution deviation is detected to meet the preset threshold condition, trigger the online fine-tuning of the initial load semantic map and the recalibration process of the nonlinear mapping network to complete the adaptive closed-loop optimization of the control strategy.

[0019] The central air conditioning source-load coordinated control method and system provided in this application have the following beneficial effects: (1) By constructing and dynamically updating the load semantic graph, the original equipment data stream is transformed into a high-level semantic representation with building physical meaning and context awareness, which effectively overcomes the inherent defects of traditional fuzzy PID control that relies on manually setting error variables and is difficult to capture complex building dynamic characteristics. On this basis, a lightweight graph neural network is used to realize real-time semantic embedding on the edge side, enabling the system to quickly identify the current functional demand state of the space (such as "rapid cooling" and "nighttime hibernation") without relying on an accurate mathematical model. This significantly improves the response sensitivity and decision rationality to changing working conditions. Especially in scenarios with frequent changes in personnel activities or seasonal changes, it avoids control delay and energy waste caused by error signal lag, and greatly enhances the environmental adaptability and stability of the control strategy.

[0020] (2) Innovatively, semantic representation is used as the core input dimension of the fuzzy controller, and adaptive PID parameters are generated by combining pre-trained mapping network and energy efficiency constraint mechanism. This completely gets rid of the limitations of manual coding of expert experience rules and fixed parameter configuration in traditional methods. Furthermore, the hierarchical fuzzy rule base is automatically generated based on semantic node attributes (such as thermal inertia and response urgency), realizing deep coupling between control strategy and building characteristics. For example, integral saturation is automatically suppressed in high thermal inertia areas, and the proportional gain is dynamically adjusted according to the demand intensity in open areas, thereby effectively balancing system energy efficiency and comfort goals while ensuring temperature control accuracy. This mechanism not only significantly improves the multi-objective collaborative optimization capability, but also makes the control logic have good interpretability and engineering adjustability, which is convenient for migration and deployment in different building types.

[0021] (3) A semantic drift detection and online graph fine-tuning mechanism is introduced to continuously monitor the distribution shift of semantic activation modes. When the evolution of usage patterns (such as work and rest adjustments or changes in functional areas) is detected, the model recalibration is automatically triggered to ensure that the control performance does not degrade during long-term operation. This effectively solves the problems of performance degradation and high maintenance costs that traditional control systems are prone to when facing dynamic changes in the building life cycle. The entire architecture does not need to perform complex instruction-level interactions with advanced optimization modules such as MPC to achieve source-load level behavior coordination, which reduces the complexity of system integration and communication overhead, and has excellent edge computing friendliness and engineering feasibility.

[0022] In summary, this solution takes the semantic essence of building load as the starting point for regulation and constructs a closed-loop adaptive control system from perception to decision-making. It realizes the formal expression of control knowledge, the contextual adaptation of the reasoning process, and the automated generation of parameter tuning. This not only significantly improves the real-time response capability, energy efficiency, and comfort level of HVAC systems, but also innovatively integrates the interpretability of artificial intelligence with the regularity of physical systems. It provides a new control paradigm with both high performance and high reliability for the field of intelligent buildings, and has broad promotion value and application prospects. Attached Figure Description

[0023] Figure 1 This is the main flowchart of a central air conditioning source-load coordinated control method; Figure 2 This is a sub-flowchart of a central air conditioning source-load coordinated control method; Figure 3 This is another sub-flowchart of a central air conditioning source-load coordinated control method; Figure 4 This is an application environment diagram of a central air conditioning source-load coordinated control method in one embodiment; Figure 5 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0025] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0026] like Figure 1 As shown, this application provides a central air conditioning source-load coordinated control method and system, specifically including: S1: Acquire the raw data stream collected by the multi-source sensors at the central air conditioning terminal. The raw data stream includes the indoor temperature setpoint and actual temperature difference, fan coil unit start / stop status, water valve opening degree and timestamp information. Perform spatiotemporal alignment and cleaning processing on the raw data stream to generate a standardized terminal energy consumption characteristic sequence.

[0027] S2: Based on the standardized end-use energy consumption feature sequence, combined with building functional zoning information and personnel activity heat map distribution, semantic annotation processing is performed on the time-series segments in the original data stream. Basic semantic units including rapid cooling demand, steady-state maintenance demand, local overcooling suppression and nighttime low-load hibernation are defined, and a basic semantic tag set with spatial weight attributes is generated.

[0028] S3: Utilize the causal relationships between the basic semantic tag set and historical operational data to construct an initial load semantic graph, where graph nodes represent the basic semantic units, and graph edges represent semantic co-occurrence relationships, temporal transition probabilities, and causal influence strength, forming a topological structure model that reflects the dynamic evolution path of building load.

[0029] S4: Dynamically embed the standardized end-point energy consumption feature sequence updated in real time into the initial load semantic map, identify the dominant semantic nodes and their neighborhood activation patterns under the current operating conditions through a lightweight graph neural network model, and calculate and output a load semantic representation vector with context awareness.

[0030] S5: Based on the load semantic representation vector and the pre-trained nonlinear mapping network, perform dynamic mapping operation of fuzzy PID control parameters. According to the thermal inertia level and response urgency in the node attributes of the initial load semantic map, calculate the direction and amplitude range of the proportional coefficient increment, integral coefficient increment and derivative coefficient increment, and generate a fuzzy PID parameter candidate set to be pruned.

[0031] S6: Based on the energy efficiency bandwidth constraints of the current chilled water system, perform safe pruning on the candidate set of fuzzy PID parameters to be pruned, eliminate parameter combinations that exceed the physical execution limit, and generate the final fuzzy PID control parameter set that meets the system stability requirements.

[0032] S7: Input the final fuzzy PID control parameter set into the fuzzy inference engine to replace the traditional fixed rule base based on error and deviation change rate. Automatically generate hierarchical fuzzy rules according to the hierarchical attributes of the nodes in the initial load semantic graph, and execute the output operation of water supply temperature setpoint adjustment and pump group frequency adjustment command.

[0033] S8: Monitor the jump status of the dominant semantic node and the distribution deviation of the neighborhood activation mode within the continuous sampling period. If the distribution deviation is detected to meet the preset threshold condition, trigger the online fine-tuning of the initial load semantic map and the recalibration process of the nonlinear mapping network to complete the adaptive closed-loop optimization of the control strategy.

[0034] S1: Acquire the raw data stream collected by multi-source sensors at the central air conditioning terminal. The raw data stream includes the indoor temperature setpoint and actual temperature difference, fan coil unit start / stop status, water valve opening degree, and timestamp information. Perform spatiotemporal alignment and cleaning processing on the raw data stream to generate a standardized terminal energy consumption characteristic sequence. Specifically, this includes: S1.1: Obtain the raw data streams collected by the smart thermostats and flow sensors deployed at each air conditioning terminal. The raw data streams include indoor temperature setpoints, actual temperature differences, fan coil unit start / stop status, water valve opening degree, and timestamp information. The raw data streams are then processed by protocol parsing and format unification to generate multi-source heterogeneous raw data packets with a unified data interface.

[0035] For the raw data streams collected by smart thermostats and flow sensors deployed at each air conditioning terminal, the input conditions are a multi-source heterogeneous data set containing indoor temperature setpoint, actual temperature difference, fan coil unit start / stop status, water valve opening degree and timestamp information.

[0036] The system parses the data communication protocols of each sensor and calls the corresponding protocol parsing module to map the private data formats output by sensors from different manufacturers and models into a unified logical structure within the system, ensuring that field names, units, and encoding methods are consistent.

[0037] Based on the parsed data fields, perform data type conversion operations to unify numerical data such as temperature, flow rate, and valve position into floating-point format, and unify start / stop status into Boolean or enumeration type encoding to ensure that subsequent processing modules can directly call them.

[0038] The parsed data is normalized by converting all temperature values ​​collected by different sensors to degrees Celsius, flow rates to cubic meters per hour, and valve positions to percentage openings, thus establishing a consistent physical quantity measurement system.

[0039] Construct a field mapping table to establish a one-to-one or one-to-many mapping relationship between the data fields of each sensor and the system's preset standard field names, and reorganize the data structure according to the mapping table to improve the consistency and scalability of the data interface.

[0040] The data encapsulation module is invoked to group the unified data fields by timestamp and device identifier, forming a multi-source heterogeneous raw data packet with time series index, multi-source identification information and standard physical quantity units.

[0041] Through the above-mentioned protocol parsing, type conversion, unit normalization, and field mapping processing methods, the collection results of the previous step are transformed into multi-source heterogeneous raw data packets that can be directly calculated by subsequent spatiotemporal alignment algorithms, thereby achieving data interface standardization and physical quantity unification.

[0042] For example, 20 smart thermostats and 20 flow sensors were deployed in a large commercial complex. The thermostats used the Modbus RTU protocol, and the flow sensors used the BACnet / IP protocol. The thermostats output the indoor temperature setpoint in Fahrenheit, and the actual temperature difference in Kelvin. The flow sensors output valve positions as integer values ​​from 0 to 255, with "ON" / "OFF" as the on / off characters. After protocol parsing, the setpoint field in the Modbus RTU register was converted to Celsius, and the formula was applied. Convert Fahrenheit to Celsius, where The original temperature value. The Kelvin value is then converted using the formula... Convert to Celsius. Convert the valve position integer value using the formula... Converted to percentage opening, where This is the original valve position value. Character-type start / stop status is mapped to Boolean 1 / 0. All values ​​are unified as floating-point numbers and associated with the temperature controller and flow sensor data by timestamp. This data is encapsulated into a multi-source heterogeneous raw data packet containing device ID, timestamp, temperature setpoint (°C), actual temperature difference (°C), valve position (%), and start / stop status (Bool). After verifying interface consistency, it is sent to the subsequent S1.2 spatiotemporal alignment processing module. This significantly improves operational stability, significantly reduces data retrieval latency, and fully guarantees interface compatibility.

[0043] S1.2: Based on the timestamp information in the multi-source heterogeneous original data packets, the missing data points are filled in using a linear interpolation algorithm and a nearest neighbor matching strategy, and the data streams with different sampling frequencies are resampled to generate a time-aligned data set with strict time synchronization characteristics.

[0044] S1.3: Based on the building spatial location coordinates of each end device in the time-series aligned data set, the physical location is mapped to a logical spatial index using geographic grid coding technology, and spatial dimension association operation is performed in combination with the device identifier in the data stream to generate a spatiotemporal coupled data matrix with a clear spatial topology relationship.

[0045] The input conditions are a time-aligned data set with strict time synchronization characteristics processed by S1.2, as well as the building spatial location coordinates and unique device identifiers corresponding to each end device.

[0046] A geogrid coding algorithm is applied to the spatial coordinates of buildings to convert two-dimensional or three-dimensional physical coordinates into fixed-precision logical spatial index codes. The coding precision is determined based on the smallest management unit of the building zoning. For example, a hierarchical grid coding is used to ensure that the location of each end device can be mapped to a unique grid index.

[0047] The obtained logical spatial index is matched with the device identifier in the time-aligned data set using key-value matching to establish a one-to-one correspondence table between the spatial index and the device identifier, which is used for subsequent spatial dimension association.

[0048] Construct a spatial topology mapping matrix based on the relationship table. The rows and columns of the matrix correspond to logical space indices, and the element values ​​are calculated from the adjacency relationship or physical distance between corresponding indices. The physical distance can be calculated using the formula: Calculation, where , Let be the planar coordinates of the i-th device.

[0049] The spatial topology mapping matrix and the time-aligned data set are matched by Cartesian product in the time dimension to form a spatiotemporal coupling record table containing the spatial location relationship of each device in each time slice.

[0050] Perform matrix reconstruction on the spatiotemporal coupling record table to generate a spatiotemporal coupling data matrix. Each row in the matrix corresponds to a device identifier and its associated timestamp, and each column corresponds to a spatial index and its distance value to adjacent spaces. The matrix elements are the combined values ​​of the device operating status data and spatial relationship weights at that moment.

[0051] The above processing method transforms the temporally aligned dataset from the previous step into a spatiotemporally coupled data matrix with a clear spatial topological relationship, realizing the structured association of the end data in the spatial dimension and providing support for subsequent anomaly detection and feature extraction.

[0052] For example, in the central air conditioning system of a large office building, the location coordinates of the terminal devices were obtained by total station calibration, with a planar coordinate accuracy of 0.01 meters. A hierarchical geogrid coding algorithm was adopted, with the first-level grid representing the floor and the second-level grid representing the planar partition, and the coding length set to 8 bits. A terminal device identifier E_102 is located at X=12.25 meters, Y=8.50 meters, and its mapped logical space index is 01020045. In the physical distance calculation, the coordinate difference between E_102 and E_105 is ΔX=1.50 meters and ΔY=2.00 meters, respectively. Substituting these values ​​into the formula: The distance was found to be 2.50 meters. This distance value was entered into the corresponding position in the spatial topology mapping matrix and paired with the operating status data of E_102 at 10:15:00 on 2024-05-08 to form a spatiotemporal coupling record table entry. After reconstruction, the matrix elements show that the device was in fan coil unit operation at this time, with a water valve opening of 45% and a spatial relationship weight of 0.40. The verification results show that the spatiotemporal coupling matrix can accurately reflect the spatial adjacency relationship and operating status of devices within the same time slice, providing a high-quality spatial data foundation for subsequent sliding window filtering and anomaly detection.

[0053] S1.4: Apply a sliding window filtering algorithm and statistical outlier detection rules to the spatiotemporal coupled data matrix to remove abnormal noise data caused by sensor failure or transmission jitter, and perform smoothing and noise reduction processing on the continuously fluctuating signal to generate a clean end-point operating status sequence with high signal-to-noise ratio.

[0054] S1.5: Based on the clean terminal operating status sequence, extract key variables such as indoor temperature setpoint and actual temperature difference, fan coil unit start / stop status, and water valve opening degree, and perform standardized normalization calculations according to the preset feature vector structure to generate a standardized terminal energy consumption feature sequence that can be directly used for subsequent semantic annotation and graph embedding.

[0055] S2: Based on the standardized end-point energy consumption characteristic sequence, combined with building functional zoning information and heat map distribution of personnel activities, semantic annotation processing is performed on the time-series segments in the original data stream. Basic semantic units are defined, including rapid cooling demand, steady-state maintenance demand, local overcooling suppression, and nighttime low-load hibernation, generating a basic semantic tag set with spatial weight attributes. Specifically, this includes: S2.1: Obtain standardized end-use energy consumption characteristic sequences and pre-set building functional zoning information. Use the time-series sliding window algorithm to fragment the temperature change rate, water valve opening fluctuation frequency and fan start-stop duty cycle in the characteristic sequence to generate a set of original data fragments containing local time-domain statistical characteristics.

[0056] The input consists of a standardized end-point energy consumption characteristic sequence generated after S1 processing and pre-set building functional zoning information. The sequence includes normalized numerical representations of indoor temperature setpoints, actual temperature differences, water valve openings, and fan coil unit start / stop status. The building functional zoning information includes the spatial zoning identifiers and usage categories corresponding to each end-point device. Based on this input, firstly, a sliding window algorithm is used to calculate the local rate of change for the time series of temperature setpoints and actual temperature differences. Within each window, the temperature rate of change is obtained using differential operations, and its sign and amplitude characteristics are recorded to form a time slice reflecting the dynamic characteristics of cooling demand. Secondly, for the water valve opening sequence, the fluctuation frequency of the opening value is statistically analyzed within the sliding window, and the total number of opening changes and the average change amplitude within the window are calculated. This is used to depict the activity level of flow regulation. Thirdly, the duty cycle of the fan coil unit start / stop status data is calculated. The ratio of the fan operating time to the total time is accumulated within the sliding window range to obtain the fan operation activity index, which serves as a periodic characteristic of the end-point air supply behavior. Subsequently, the three indicators—temperature change rate, water valve opening fluctuation frequency, and fan duty cycle—are combined according to window timestamps into raw data fragments containing local time-domain statistical characteristics. Each fragment includes building functional zoning location information, providing a foundation for subsequent spatial weighted matching using heatmaps of human activity. Through the above processing method, the results of the previous step are transformed into a set of raw data fragments capable of carrying multiple energy consumption behavior characteristics, enabling the simultaneous extraction of local behavioral patterns and spatial positioning of end-point operating status signals.

[0057] For example, in a central air conditioning system of a well-defined office building, the sampling period for collecting standardized terminal energy consumption characteristic sequences is 60 seconds. Pre-set building functional zoning information includes four types of zones: offices, meeting rooms, corridors, and leisure areas. In the temperature change rate calculation, the sliding window length is set to 10 minutes and the window step size to 2 minutes. In the sequence of an office zone, the temperature setpoint within a certain window is... ℃, the actual temperature difference shows that from ℃ dropped to The rate of change of ℃ is obtained by difference calculation. The temperature fluctuation rate (℃ / second) is negative, indicating a cooling trend. The frequency of water valve opening fluctuations within this window is [number] times. The average fluctuation was [number] times. % opening degree; the fan coil unit is in working state during start / stop. Minutes, Close The formula for calculating the duty cycle is: minutes The result is 0.8. The above three indicators, along with the window start and end timestamps and the corresponding office space identifiers, are encapsulated into a raw data fragment. This fragment will be used in S2.2 to calculate the spatial weight in conjunction with the personnel activity heatmap. For the conference room scenario, the temperature change rate may be positive (heating trend), the water valve opening frequency is zero, and the fan duty cycle is extremely low, indicating a low load state. Such fragments are more likely to be classified into steady-state maintenance or nighttime low-load hibernation semantic units in subsequent semantic discrimination. Ultimately, after semantic annotation, these fragments can support the load semantic map in modeling the cooling demand path of different spaces, significantly improving the adaptability and robustness of the control strategy.

[0058] S2.2: Based on the original set of data fragments and the real-time heat map distribution data of personnel activities, perform multi-source data spatiotemporal correlation matching operation, calculate the personnel density weight coefficient and thermal disturbance intensity index of the area where each data fragment is located, so as to generate a weighted time-series feature vector group with spatial location attributes.

[0059] S2.3: Using the weighted time-series feature vector group, the predefined semantic rule engine is called to perform logical discrimination processing on the temperature deviation trend and flow response mode in the vector, identify and mark four basic semantic units: rapid cooling demand, steady-state maintenance demand, local supercooling suppression and nighttime low-load hibernation, so as to generate a preliminary discrete semantic label sequence.

[0060] Based on the weighted time-series feature vector set output from S2.2, a predefined semantic rule engine is invoked to perform logical recognition according to a two-dimensional discrimination framework of temperature deviation trend and flow response mode. The temperature deviation trend index and flow response mode features are input into the temperature difference channel and flow channel of the rule engine, respectively. A trend direction determination operation is performed on the temperature difference channel, using a sign function to identify the increase, decrease, or stable state of temperature difference change. A flow change rate threshold comparison is performed on the flow channel to determine whether it is in a high response, low response, or no response range. For the temperature difference direction determination result and the flow range determination result, combined judgment entries from the rule base are invoked to map different combinations to specific semantic unit categories. For example, a significant decrease in temperature difference and a high flow change rate are mapped to a rapid cooling demand, while a stable temperature difference and a low flow change rate are mapped to a steady-state maintenance demand. For segments with multiple feature combinations that cross-match, priority conflict resolution logic is executed to ensure that the same time-series segment corresponds to a unique semantic category label. The determined semantic categories are encapsulated into a sequence structure to form a preliminary discrete semantic label sequence as output. Through the above logical discrimination processing method, the weighted temporal feature vector group of the previous step is transformed into semantic label data that can be used for subsequent membership degree calculation and graph construction, so as to achieve accurate semantic classification of the terminal operating conditions.

[0061] For example, in a large commercial building, for the weighted time-series feature vector group of the east wing conference area, the temperature deviation trend value is calculated to be +0.8℃ / h, and the flow response mode change rate is 0.25 openings / h. The rule engine's temperature difference channel determines that the temperature deviation is increasing, and the flow channel determines that the flow is in the medium response range. Based on the rule base combination entries, the combination of increasing temperature difference and medium response is mapped to a local overcooling suppression semantic unit. During the judgment process, the sign function calculates the temperature trend direction as +1, the threshold comparison identifies the flow range category code as 2, and the combination lookup table matches the semantic code 3. In the output discrete semantic label sequence, this time segment is assigned a local overcooling suppression label. Subsequently, S2.4 calculates the spatial range and temporal duration weight of this label. The verification results show that the matching degree of the cooling capacity adjustment strategy of this conference area during this period is significantly improved, avoiding energy waste caused by temperature overshoot.

[0062] S2.4: Based on the discrete semantic tag sequence and the personnel density weight coefficient, the confidence of each semantic unit on the time axis is quantified by the fuzzy membership function to determine the spatial scope and time duration weight of each semantic tag, so as to generate basic semantic annotation results with continuous numerical attributes.

[0063] S2.5: The basic semantic annotation results with continuous numerical attributes are subjected to structured encapsulation processing, and the semantic unit type, spatial weight attribute and timestamp information are integrated into a unified data structure to finally generate a basic semantic tag set with spatial weight attribute for driving the construction of the load semantic graph.

[0064] like Figure 2 As shown, S3: Utilizing the causal relationships between the basic semantic tag set and historical operational data, an initial load semantic graph is constructed. Graph nodes represent the basic semantic units, and graph edges represent semantic co-occurrence relationships, temporal transition probabilities, and causal influence strength, forming a topological model reflecting the dynamic evolution path of building loads. Specifically, this includes: S3.1: Obtain the basic semantic tag set with spatial weight attributes and the cleaned historical running data. Perform sliding window statistical processing on the temporal occurrence sequence of each basic semantic unit in the historical running data to calculate the co-occurrence frequency of any two basic semantic units in the same time window and generate a basic semantic co-occurrence frequency matrix.

[0065] The process involves acquiring a basic semantic tag set containing spatial weight attributes and cleaned historical operational data. The timestamps and spatial indices of each semantic unit in the basic semantic tag set are analyzed to establish a temporal position mapping for these semantic units. Based on this mapping, sensor state sequences corresponding to each semantic unit are extracted from the historical operational data, including temperature difference changes, water valve operation states, and fan coil unit start / stop records, forming a temporal occurrence sequence set for each semantic unit. For each semantic unit's temporal occurrence sequence, a sliding window statistical algorithm with a fixed window length is applied to determine if two semantic units are simultaneously present within each time window. The co-occurrence count within the window is counted and stored in a two-dimensional index structure. A matrix accumulation method is used to summarize the co-occurrence counts within all time windows into a preliminary co-occurrence frequency matrix, where the matrix rows and columns represent semantic unit pairs, and the matrix element values ​​represent the co-occurrence frequency within the time window. Spatial weight correction is performed on the preliminary co-occurrence frequency matrix, using the spatial weight attributes from the semantic tag set as coefficients to perform weighted multiplication operations on the matrix element values, thereby improving the true impact of co-occurrence frequency in different areas of the building. By using the weighted and corrected matrix, the co-occurrence frequency of any two basic semantic units within the same time window is accurately calculated, generating a basic semantic co-occurrence frequency matrix, which provides input for subsequent conditional probability estimation and causal testing.

[0066] For example, for a large commercial building operating under high load conditions in summer, a sliding window length of 15 minutes and a window step of 5 minutes are selected. A tag set containing four basic semantic units is used: rapid cooling demand, steady-state maintenance demand, local overcooling suppression, and nighttime low-load hibernation. Based on the tag set mapping, the corresponding time-series occurrence sequences are extracted from historical operating data, and the co-occurrence of the above semantic units in each time window is statistically analyzed. For example, in the co-occurrence statistics of rapid cooling demand and steady-state maintenance demand, if both occur simultaneously 3 times within a certain window, the matrix element value is 3. Spatial weight attributes are applied to this value, such as the weight coefficient of the area where rapid cooling demand is located. The weighting coefficient for the region where steady-state maintenance demand is located is The corrected co-occurrence frequency is obtained through the formula. Calculate and obtain the correction value. After traversing all windows and summing the matrices, the generated basic semantic co-occurrence frequency matrix has an element of 2.88 corresponding to the rapid cooling requirement and the steady-state maintenance requirement. The remaining elements are calculated using the same method. This matrix serves as the core input for conditional probability calculation in subsequent steps, effectively improving the regional adaptability and temporal accuracy of semantic relationship modeling.

[0067] S3.2: Based on the co-occurrence frequency matrix of the basic semantics, the conditional probability estimation algorithm is used to quantify the successive relationship between basic semantic units to determine the conditional probability value of the transition from the previous basic semantic unit to the next basic semantic unit, and generate a set of basic semantic temporal transition probabilities.

[0068] Obtain the basic semantic co-occurrence frequency matrix generated by S3.1. Construct a conditional probability calculation task based on the statistical values ​​of the co-occurrence of any two basic semantic units within the same time window. Represent each row of the matrix as the frequency set of the preceding semantic unit, and each column as the frequency set of the subsequent semantic unit. Normalize the results to obtain the total frequency of the preceding semantic units as the denominator. Employ a conditional probability estimation algorithm, using the co-occurrence frequency of each pair of semantic units as the numerator and performing a fractional operation with the sum of the preceding frequencies to form the initial value of the time transition probability. The conditional probability formula is expressed as: in It is a frequency statistics function. and These represent the preceding and succeeding basic semantic units, with the numerator representing the co-occurrence frequency of the two units and the denominator representing the total occurrence frequency of the preceding semantic unit. A probability boundary check is performed on the initial values ​​to eliminate temporal relationships with probability values ​​less than the system noise threshold, thus reducing the interference of anomalous short-term co-occurrences on the transition probability model. The probability values ​​that pass the check are combined with the corresponding semantic unit pairs to form a basic semantic temporal transition probability set, providing quantitative input for subsequent causal significance verification. Using a conditional probability estimation algorithm, the co-occurrence frequency data from the previous step is transformed into a temporal transition probability index reflecting the strength of the successive relationship, achieving dynamic evolution and quantification of semantic relationships.

[0069] For example, in the summer operation data of a large office building, the co-occurrence frequency of rapid cooling demand and steady-state maintenance demand in the basic semantic co-occurrence frequency matrix is ​​120 times, and the total occurrence frequency of rapid cooling demand is 200 times. Applying the conditional probability formula, the numerator is... The denominator is The conditional probability is calculated. After probability boundary checks, the system noise threshold was set to 0.05, and this value is retained in the set. The co-occurrence frequency of local overcooling suppression and steady-state maintenance requirements within the same matrix is ​​30 times, and the total occurrence frequency of local overcooling suppression is 500 times, with a conditional probability of... The values ​​were retained after passing the test. The final set contains multiple pairs of semantic units and their corresponding probabilities, such as (rapid cooling → steady-state maintenance) = 0.6, (local overcooling inhibition → steady-state maintenance) = 0.06, etc. In the subsequent Granger causality test, these probabilities are used to evaluate the significance of causal pointing relationships and ensure that the model accurately captures the actual dynamic laws of semantic transfer.

[0070] S3.3: Based on the set of basic semantic temporal transition probabilities and the temperature difference change rate and valve action delay data in the historical operation data, the Granger causality test method is used to verify the significance of the driving and driven relationship between basic semantic units, so as to extract the statistically significant causal pointing relationship and generate a list of basic semantic causal influence intensity.

[0071] S3.4: Using the basic semantic units as a set of vertices, and combining the basic semantic co-occurrence frequency matrix, the basic semantic temporal transition probability set, and the basic semantic causal influence intensity list as edge attribute data, perform a topology assembly operation of a directed weighted graph to establish multidimensional connections between nodes and generate an initial load semantic graph topology.

[0072] Using the extracted basic semantic units as the vertex set, the basic semantic co-occurrence frequency matrix, the basic semantic temporal transition probability set, and the basic semantic causal influence strength list are read as edge attribute inputs. The vertex set is initialized with directed weighted graph topology construction, and each semantic unit is assigned a unique node index to establish node reference relationships.

[0073] For each element value of the co-occurrence frequency matrix, the edge weight calculation module is called to normalize the co-occurrence frequency to the interval [0,1] to form the co-occurrence intensity coefficient, and a preliminary undirected edge structure is constructed for the corresponding node pairs.

[0074] Based on the conditional probability values ​​in the temporal transition probability set, the direction assignment algorithm is called to transform the initial undirected edge into a directed edge, and the conditional probability values ​​are written into the temporal weight attribute of the edge, so that the directed edge can reflect the possibility of transition from the source node to the target node.

[0075] The significance test results are read from the list of causal influence strengths. The multidimensional weight fusion function is then called to weight and synthesize the causal strength value with the co-occurrence strength coefficient and the time series weight to form a comprehensive edge weight. ,in , , For preset weighting coefficients, The co-occurrence intensity coefficient, For time series weights, This represents the causal strength value.

[0076] For all node sets and the combined edge weights, construct the graph adjacency list of the directed weighted graph, write the edge attributes into the adjacency list entries and create an index for subsequent graph algorithm calls.

[0077] Based on the constructed adjacency list, the topology verification mechanism is invoked to check for the existence of isolated nodes, loops, and redundant edges, and edge filling or deletion operations are performed according to the rules to ensure the integrity and usability of the graph structure.

[0078] Through the above-mentioned multi-dimensional weight fusion and topology assembly processing, the relational data of the previous step is transformed into an initial load semantic map topology with directionality, weight attributes and causal relationships, thereby realizing a structured expression of the dynamic evolution path of building load.

[0079] For example, in the central air conditioning system of an office building, there are four basic semantic units, with indices of 0 (rapid cooling requirement), 1 (steady-state maintenance requirement), 2 (local overcooling suppression), and 3 (nighttime low-load hibernation); the co-occurrence frequency matrix is ​​obtained after normalization. Equal coefficients; in the time transition probability set, the conditional probability of moving from node 0 to node 1 is... The conditional probability of the path from node 1 to node 2 is: The strength value of node 0 driving node 1 in the causal influence strength list is... The weighting coefficients are set to... , , The comprehensive edge weight calculation is as follows The output result is In the generated topology, an edge with a direction of 0→1 and a weight of 0.805 is established between node 0 and node 1. After checking by the topology verification mechanism, this graph has no isolated nodes, has a cycle count of 1, and can be retained for dynamic evolution analysis. Finally, an initial semantic graph topology structure adapted to the dynamic load pattern of the building is formed, which can be directly used for subsequent graph neural network embedding and parameter mapping models.

[0080] S3.5: Based on the initial load semantic graph topology, add thermal inertia level attributes and response urgency attributes to each graph node and perform metadata enhancement processing to improve the state description dimension of the node, and generate the final initial load semantic graph containing complete node attributes and edge weights.

[0081] Based on the initial load semantic graph topology, the identifier of each basic semantic unit in the node set and its associated edge weight data are read as input objects for metadata enhancement processing.

[0082] The system calls upon the building physics database to calculate the transient temperature rise response time of the area based on the heat transfer coefficient of the building envelope, building volume, and material heat capacity of the region mapped by the node, and assigns a thermal inertia level attribute accordingly. The thermal inertia level is classified according to the following formula: in, For material density, For specific heat capacity, For the volume of the region, The heat transfer coefficient is... is the external surface area.

[0083] Read the historical load fluctuation frequency and causal intensity list associated with the nodes, calculate the average delay of demand changes using time series event interval statistics, and classify them into different response urgency levels; the calculation of response urgency can be expressed as: in Let be the interval between adjacent demand events of the i-th demand event, and n be the total number of demand event intervals within the statistical window.

[0084] The thermal inertia level attribute and response urgency attribute are appended to the metadata structure of the corresponding node in the form of key-value pairs, maintaining a consistent reference format between node attribute data and edge weight data.

[0085] Perform consistency checks on the enhanced node metadata to ensure that all nodes have complete thermal inertia level and response urgency fields and match the semantic identifiers of the existing graph topology.

[0086] By enhancing node attributes, the topology results from the previous step are transformed into complete semantic graph node data with physical characteristics and dynamic response features, enabling the load semantic graph to have higher context awareness and scene adaptation capabilities when embedded into subsequent graph neural networks.

[0087] For example, in a large commercial complex, the node "Meeting ends → Personnel leave → Localized overcooling suppression" corresponds to a 200m² conference room, and the density of the insulation material in the building envelope is... kg / m³, specific heat capacity kJ / (kg·K), volume m³, heat transfer coefficient W / (m²·K), external surface area m². Substitute into the formula to calculate the thermal inertia level. = It is classified as a high thermal inertia level. The average event interval for historical load changes at this node is... Minutes, Response Urgency Calculation = min - ¹, classified as low urgency level. The metadata-enhanced nodes, when embedded in the graph neural network, can accurately provide regional thermal buffering capacity and response time parameters, enabling the MPC and fuzzy PID co-optimization strategy to significantly improve the stability and matching degree of water supply temperature setting adjustment in this scenario, and reduce the risk of cold source overshoot.

[0088] like Figure 3 As shown, S4 involves dynamically embedding the real-time updated standardized end-point energy consumption feature sequence into the initial load semantic map, identifying the dominant semantic nodes and their neighborhood activation patterns under the current operating conditions using a lightweight graph neural network model, and calculating and outputting a load semantic representation vector with context-aware capabilities. Specifically, this includes: S4.1: Obtain the standardized end-use energy feature sequence generated by spatiotemporal alignment and cleaning, perform node matching calculation on the time segments in the sequence based on the definition rules of the basic semantic tag set, and transform the discrete multi-source sensor data into initial semantic node embedding vectors to establish a preliminary mapping relationship between physical measurement data and the spectral node space.

[0089] The standardized end-point energy consumption feature sequence generated by spatiotemporal alignment and cleaning is obtained as the input object. Physical measurement data such as indoor temperature setpoint, actual temperature difference, water valve opening degree and fan coil start-stop status of each time segment in the sequence are read. The definition rule library in the basic semantic tag set is called to perform semantic condition matching calculation on each segment to determine its corresponding semantic unit type.

[0090] For the temperature difference parameter in the matching determination, the normalized rate of change of temperature difference is used as the numerical input, and the calculation formula is as follows: ,in For segment temperature difference, The mean of the entire sequence. The standard deviation is used to quantify the contribution of temperature difference to semantic matching.

[0091] For the water valve opening parameter, a piecewise linear mapping is performed, projecting the opening percentage value to the corresponding membership value in the interval [0,1]. The mapping relationship is as follows: Definition, where This represents the opening value of the water valve.

[0092] Multi-source physical measurement data that has undergone semantic matching are indexed and bound according to the unique identifier of the node in the basic semantic label set, and a key-value correspondence between the original physical quantity and the semantic node is established.

[0093] The binding results are processed by feature vectorization, combining each physical quantity into node embedding vector elements, following the... Structure, in which To normalize the temperature difference, To normalize the water valve opening, This is a binary code for the start-stop status of the wind turbine.

[0094] By mapping the node embedding vector set one-to-one with the node space of the initial load semantic graph, a preliminary mapping relationship between physical measurement data and semantic node space is achieved.

[0095] Through the above processing method, the standardized end-point energy consumption feature sequence of the previous step is transformed into a node embedding vector for the initial load semantic map, realizing a physical data input format with semantic attributes, and providing an efficient and interpretable input data structure for graph neural networks to perform dynamic working condition identification.

[0096] For example, in the central air conditioning system of a large commercial building, a sequence of end-point energy consumption characteristics, after spatiotemporal alignment and cleaning, is obtained. This sequence includes an indoor temperature setpoint of 23.0℃, an actual temperature difference of 2.5℃, a water valve opening of 60%, and a fan coil unit on / off status, collected every 5 seconds. The basic semantic tag definition rules are used to determine that this segment belongs to "rapid cooling demand," and the normalized temperature difference change rate is calculated as follows: =1.4, the membership value of the water valve opening is calculated as follows: =0.6, and the fan start / stop status code value is 1. The three types of parameters are combined into a node embedding vector [1.4, 0.6, 1], which is then bound to the corresponding "rapid cooling demand" node in the semantic graph, completing the initial mapping from physical measurement data to the semantic node space. In subsequent operating condition identification, this embedding vector can significantly improve the accuracy of identifying the dominant semantic node and ensure the model's response speed and robustness under high thermal disturbance scenarios.

[0097] S4.2: Using the initial semantic node embedding vector as input, the lightweight graph neural network model deployed on the edge side is invoked to perform message passing and aggregation operations. The hidden state of the node is updated according to the edge weight attributes defined in the initial load semantic graph, and a dynamic node feature representation that integrates the temporal transition probability and the causal influence intensity is generated.

[0098] The process begins by obtaining initial semantic node embedding vectors as input. A lightweight graph neural network model deployed at the edge is then invoked. Message passing is triggered using the adjacency matrix and edge weights, propagating the embedding vectors of each node along the edges of the graph to form a neighborhood message set. This neighborhood message set is then aggregated by weighting the sum based on the temporal transition probabilities and causal influence strength values ​​corresponding to the edge weights in the graph, preserving the temporal and causal characteristics of load evolution during the aggregation process. The aggregated neighborhood messages are then fused with the current node embedding vector, mapped to the updated hidden state space using a nonlinear activation function. This introduces both temporal and causal information from the context into the node representation. The updated hidden state is then normalized to eliminate the impact of input distribution differences between different nodes on the stability of subsequent calculations, ensuring the comparability of dynamic feature representations under different operating conditions. This process transforms the initial semantic embedding vectors from the previous step into dynamic node feature representations that integrate temporal transition probabilities and causal influence strengths, enabling context-aware modeling of the building load evolution path.

[0099] For example, in the central air conditioning control system of an office building, the standardized end-point energy consumption feature sequence collected by the edge gateway is processed by S4.1 to obtain a 12-dimensional initial semantic node embedding vector. Each dimension corresponds to features such as temperature difference change rate, water valve opening, and fan duty cycle. The lightweight graph neural network sets the number of hidden layer nodes to 64 and adopts a three-stage mechanism of message passing, aggregation, and update. During the message passing process, for any node v, the formula for calculating its neighborhood message set M(v) is: in This represents the edge weight from node u to node v. Let represent the embedding vector of node u at time t. Let v be the set of neighbors of node v.

[0100] Introducing time transition probability in the aggregation phase Using the causal influence strength c(u,v) as a multiplicative factor, the weight update value is calculated: The update phase uses a non-linear activation function σ, such as ReLU, to map the fused information. In the formula, This represents the embedding vector of node u at time t+1.

[0101] In this scenario, edge weights are provided by the co-occurrence frequency matrix in the initial load semantic graph, with temporal transition probabilities ranging from 0.2 to 0.8 and causal influence strength ranging from 0.3 to 1.0. After three iterations, the dynamic feature representation of each node converges stably in the 64-dimensional space, with significant differences between nodes. The causal chain of the dominant node can be accurately identified, ultimately achieving high-precision perception of the dominant semantic node and its neighborhood activation patterns, providing reliable input for subsequent S4.3 attention score calculation.

[0102] S4.3: Based on the dynamic node feature representation, perform attention mechanism scoring calculation, quantify the activation salience of each semantic node under the current operating conditions, select the node with the highest activation score as the dominant semantic node, and extract the set of neighboring nodes within its preset hop count range to construct a local subgraph topology.

[0103] The dominant semantic node is the graph node with the highest activation score selected at a specific time after performing an attention mechanism scoring calculation on the dynamic node feature representation. The dominant semantic node represents the most core and influential semantic unit of the current end-point energy consumption feature sequence in the load semantic graph, and its state and attributes dominate the core semantic pattern of the current building load behavior.

[0104] S4.4: Perform feature-weighted fusion processing on the set of neighboring nodes in the local subgraph topology, combine the attribute information of the dominant semantic node with the spatial distribution characteristics of the neighborhood activation mode, compress the high-dimensional topological information through the graph readout function, and output a load semantic representation vector that reflects the dynamic evolution path of the building load.

[0105] The neighborhood activation pattern refers to the feature distribution and interaction state of the set of neighboring nodes within a preset hop range centered on the dominant semantic node. This neighborhood activation pattern is characterized by spatial distribution characteristics and feature weighted fusion processing, reflecting the dynamic pattern in the local subgraph topology where related semantic units around the dominant node are collaboratively activated or influenced.

[0106] The local subgraph topology obtained from S4.3 is used as the input object. The set of neighboring nodes and the corresponding node feature vectors are read, and a neighborhood feature index table is built in memory for subsequent fusion calls.

[0107] A weighted coefficient matrix is ​​constructed based on the spatial location attributes and activation saliency scores of the neighborhood node set. The spatial location weight is derived from the spatial distribution density calculation results of the graph nodes, and the activation saliency weight is derived from the attention mechanism score. The two types of weights are then normalized by multiplication to form a superposition coefficient.

[0108] The feature weighted fusion algorithm is invoked to multiply the dynamic feature representations of each neighboring node element by element with the weighting coefficient matrix and then perform an accumulation and summation operation to generate a fused neighborhood feature aggregation vector, so as to preserve the synergistic effect of spatial distribution pattern and saliency information.

[0109] Read the node attributes such as thermal inertia level and response urgency of the dominant semantic node, concatenate the attribute vector with the neighborhood feature aggregation vector using the attribute injection method, and perform fully connected layer mapping to encode the concatenated high-dimensional vector into a compact semantic state representation.

[0110] The graph readout function is called to perform feature compression on the encoded high-dimensional semantic state representation. Using a joint strategy of global average pooling and max pooling, the long vector is mapped into a fixed-dimensional load semantic representation vector to facilitate input processing of the subsequent nonlinear mapping network.

[0111] The load semantic representation vector is a comprehensive vector generated by integrating the attribute information of the dominant semantic node and the spatial distribution characteristics of its neighborhood activation patterns, after feature weighted fusion and high-dimensional topological information compression. This load semantic representation vector is a holistic, structured mathematical expression of the current building load state at the semantic level, and can be used for downstream load forecasting, anomaly diagnosis, or optimization control tasks.

[0112] By using neighborhood feature weighted fusion and graph readout function processing, the results of the previous step are transformed into high semantic density representation data containing the dynamic evolution path of building load, thereby enhancing the context awareness capability and providing a stable and global perspective input for subsequent dynamic mapping of fuzzy PID parameters.

[0113] For example, in the central air conditioning system of a large commercial complex, the local subgraph topology includes one dominant semantic node and five neighboring nodes. The dominant node has a thermal inertia level of 3 and a response urgency of 0.8. The spatial position weights of the neighboring nodes are 0.6, 0.5, 0.7, 0.4, and 0.9, respectively, and their activation saliency scores are 0.75, 0.65, 0.80, 0.55, and 0.95, respectively. When constructing the superposition coefficient, multiplicative normalization is used, and the calculation formula is as follows: in, Spatial location weights, To activate the saliency score. For example, the superposition coefficient of the first neighboring node is... = The dynamic feature representations of each neighboring node (unified as 128-dimensional vectors with standardized values) are multiplied by their corresponding superposition coefficients and accumulated to obtain the fused aggregate vector. The dominant node attribute vector (2-dimensional) is concatenated with the aggregate vector to form a 130-dimensional high-dimensional state representation, which is then compressed to 64 dimensions via a fully connected layer. Finally, a graph readout function is called to perform global average pooling and max pooling, mapping the result to a 32-dimensional load semantic representation vector. This representation vector exhibits significantly improved parameter prediction stability in the subsequent fuzzy PID mapping network, resulting in more accurate control responses for the water supply temperature setpoint and pump frequency regulation, and demonstrating global coordination characteristics.

[0114] S5: Based on the load semantic representation vector and the pre-trained nonlinear mapping network, perform dynamic mapping operation of fuzzy PID control parameters. According to the thermal inertia level and response urgency in the node attributes of the initial load semantic map, calculate the direction and amplitude range of the proportional coefficient increment, integral coefficient increment, and derivative coefficient increment, generating a candidate set of fuzzy PID parameters to be pruned. Specifically, this includes: S5.1: Obtain the load semantic representation vector with context awareness output from the previous steps, as well as the graph node attribute data containing thermal inertia level and response urgency attributes. Construct a nonlinear mapping network model using a multilayer perceptron architecture, and perform high-dimensional feature space projection processing on the load semantic representation vector to generate the hidden layer activation state matrix, thus establishing the nonlinear correlation basis between semantic features and control parameters.

[0115] Obtain the load semantic representation vector generated by the previous steps and the metadata of the graph nodes with numerical attributes such as thermal inertia level and response urgency. Call the multilayer perceptron to construct the input layer of the nonlinear mapping network. After normalization, the input vector is loaded into the network to ensure that the distribution of features with different dimensions is consistent.

[0116] After normalization is completed in the input layer, the weight matrix multiplication and bias vector addition of the first hidden layer are performed to project the semantic features to a high-dimensional feature space. The ReLU function is used to achieve nonlinear activation to enhance the feature representation capability and retain the high-order components of the nonlinear response features in the load semantics.

[0117] The output of the first hidden layer is used as the input of the second hidden layer. Batch normalization and weight update calculations are performed. The hyperbolic tangent (tanh) activation function is used to compress the numerical range, form a stable feature distribution and reduce the risk of gradient vanishing. In this way, the nonlinear mapping relationship between building load semantics and PID control parameters is extracted in layers.

[0118] Thermal inertia level and response urgency are introduced as additional inputs in the second hidden layer. They are directly concatenated to the hidden layer feature vector through the Concatenate operation to form a high-dimensional feature group that integrates node attributes, so as to achieve targeted processing of the differences in thermal response characteristics in different regions.

[0119] The fused high-dimensional feature set is input into the third hidden layer, and weight matrix multiplication and bias addition operations are performed. The non-negativity of the output is guaranteed by the Softplus function, and the hidden layer activation state matrix is ​​generated. This matrix establishes the nonlinear correlation basis for the parameter calculation of the three control channels of proportional, integral and differential.

[0120] By using the multilayer perceptron processing method described above, the semantic representation vector from the previous step is transformed into a hidden layer activation state matrix, thereby achieving stable and scalable nonlinear correlation modeling between semantic features and control parameters.

[0121] For example, in the central air conditioning system of a building, the input load semantic representation vector is 64-dimensional, with a thermal inertia level of 1.8 and a response urgency of 0.75. The hidden layer is configured as a three-layer structure: the first hidden layer has 128 nodes, with weight matrix elements ranging from -0.05 to 0.05, and uses ReLU activation; the second hidden layer has 256 nodes, with batch normalization parameters having a mean of 0.0 and a variance of 1.0, and uses tanh activation; the third hidden layer has 3 nodes and uses Softplus activation. In the first hidden layer, a projection operation is performed, i.e., the input vector is multiplied by the weight matrix and a bias vector is added to obtain the projected feature vector. ,in This is the normalized semantic representation vector of the load. This is the weight matrix. The bias vector is used; after ReLU activation, the first layer output is obtained. The second layer concatenates the first layer output with the thermal inertia level and response urgency to form a feature set of length 66, and then performs... ,in For thermal inertia level, To address the urgency level, tanh compression is applied, resulting in output values ​​in the range (-1, 1). The third layer calculates the hidden layer activation state matrix. ,in and For the third layer weights and biases, Softplus ensures the output is non-negative. The execution results output values ​​of 2.3, 1.1, and 0.9 in the three channels, respectively, serving as the basis for calculating the subsequent parameter offsets in the proportional, integral, and derivative channels. The verification results show that the hidden layer activation state matrix generated by the network in this scenario can significantly improve the stability and interpretability of the parameter dynamic mapping.

[0122] S5.2: Based on the hidden layer activation state matrix and the pre-trained weight coefficient matrix, perform a weighted summation operation on the fully connected layer to transform the abstract semantic features into preliminary parameter adjustment trend signals, and output the original parameter offset sequence containing the proportional channel, integral channel and differential channel, thus completing the preliminary numerical mapping from the semantic space to the parameter space.

[0123] S5.3: Read the original parameter offset sequence and combine it with the thermal inertia level in the current map node attributes. Use an adaptive gain scaling algorithm to perform amplitude correction processing on the original parameter offset sequence to match the thermal response characteristics of different building areas, generate an intermediate parameter adjustment vector with thermal inertia adaptability, and eliminate the risk of control overshoot or hysteresis caused by thermal capacity differences.

[0124] Obtain the original parameter offset sequences of the proportional, integral, and differential channels, and use the thermal inertia level values ​​in the associated graph node attribute data as the basis for gain adjustment.

[0125] The thermal inertia level values ​​are analyzed and transformed into a gain adjustment coefficient matrix that can be used to quantify the differences in thermal response speed, ensuring that the thermal capacity differences in different building areas are explicitly mapped to the parameter adjustment process.

[0126] An adaptive gain scaling algorithm is used to perform channel-by-channel multiplication on the original parameter offset sequence and the gain adjustment coefficient matrix to form a gain correction vector that matches different thermal response characteristics.

[0127] When performing amplitude correction, the constraint function is called to limit the gain correction vector to avoid overshoot in the low thermal inertia region and hysteresis in the high thermal inertia region. The limiting threshold is set according to the physical response limit of the device.

[0128] The values ​​of each channel after adaptive gain scaling and limiting are merged to form an intermediate parameter adjustment vector. This vector maintains the same dimensional structure as the original parameter offset sequence, but the parameter amplitude has been adapted to the regional thermal inertia characteristics.

[0129] By using an adaptive gain scaling algorithm, the original parameter offset sequence from the previous step is transformed into an intermediate parameter adjustment vector that has thermal inertia adaptability and can eliminate the risk of overshoot or hysteresis, thereby enabling the controller to achieve steady-state accurate response in multi-regional environments.

[0130] For example, in a well-defined commercial building, the original parameter offset of the proportional channel is... The points channel is Differential channels are The corresponding thermal inertia levels are high, medium, and low, with the high thermal inertia level mapped to the gain adjustment coefficient. Medium level mapping to Low-level mapping to An adaptive gain scaling algorithm is used to perform product operations on the proportional channels. get Perform product operation on the integral channel. get Perform product operation on the differential channel get The amplitude limit constraint is set to the maximum of the proportional channel. Maximum points channel The largest differential channel The intermediate parameter adjustment vector after limiting is a proportional vector. ,integral ,differential When this adjustment vector was applied to the control of the central air conditioning chilled water system, the verification results showed that the temperature response in the high thermal inertia region was more stable, the cooling rate in the low thermal inertia region was significantly improved, and no system oscillation or equipment overload occurred.

[0131] S5.4: Based on the intermediate parameter adjustment vector and the response urgency attribute in the graph node attributes, the direction determination logic function is used to calculate the change trend of each control channel, determine the incremental direction of the proportional coefficient, the incremental direction of the integral coefficient and the incremental direction of the derivative coefficient, and calculate the corresponding amplitude range in combination with the preset step size limit threshold, and generate a fuzzy PID parameter increment set with directional constraints and amplitude boundaries.

[0132] Based on the intermediate parameter adjustment vector and the response urgency attribute in the graph node attributes, an input matrix for the direction determination logic function is constructed. The correction values ​​of the three control channels are paired one-to-one with their corresponding urgency levels to achieve data-driven direction determination preparation. The input matrix is ​​then fed into the direction determination logic function module. This module uses a combination of symbolic analysis and threshold comparison. When the product of the correction value and the urgency level is greater than zero and the urgency level exceeds a set threshold, it is determined as a positive increment; otherwise, it is determined as a negative increment, and the determination result generates a direction identifier vector. The direction identifier vector and the intermediate parameter adjustment vector are then used for element-wise pairing to obtain the preliminary increment vector after direction adjustment, which is then fed into the amplitude constraint calculation module. The amplitude constraint calculation module performs amplitude clipping calculations for each control channel based on a preset step size limit threshold set, using the formula: in, This is the increment of the control parameter for the proportional channel. The proportional channel step size limit threshold is used, and the `clip` function is the amplitude clipping function, used to truncate the incremental amplitude within the legal range. The same clipping calculation is performed on the proportional, integral, and derivative channels respectively, generating a set of fuzzy PID parameter increments with directional constraints and amplitude boundaries. Through this processing method, the thermal inertia adaptation increment vector from the previous step is transformed into parameter adjustment data with clearly defined direction and limited amplitude, achieving dual constraints on the fuzzy PID parameter increments in terms of control trend and amplitude safety.

[0133] For example, in the central air conditioning system of a large commercial building, the intermediate parameter adjustment vectors obtained by S5.3 are ΔKp=0.25, ΔKi=0.05, and ΔKd=-0.02, respectively. The response urgency level in the graph node attributes is proportional channel 3, integral channel 1, and differential channel 2. The input matrix of the direction determination logic function is [[0.25,3],[0.05,1],[-0.02,2]]. The symbol analysis determines that the proportional channel meets the positive increment condition, the integral channel meets the positive increment condition, and the differential channel does not meet the positive increment condition, so it is marked as a negative increment. The direction identification vector is [+1,+1,-1], which, after being matched with the adjustment vector, yields the initial increment vector [0.25,0.05,0.02]. The preset step size limit threshold set is Lp=0.2, Li=0.08, Ld=0.05. Executing the clip function prunes the parameters, truncating the proportional channel increment to 0.2, keeping the integral channel increment at 0.05, and the derivative channel increment at 0.02. The resulting fuzzy PID parameter increment set is [0.2, 0.05, 0.02]. This set ensures both the correctness of the response trend and that the increment magnitude remains within the system's safe operating range. In practical applications, it can significantly improve the stability and tracking accuracy of the central air conditioning source-load coordinated control during periods of rapid load fluctuations.

[0134] S5.5: Integrate the proportional coefficient increment, integral coefficient increment, and derivative coefficient increment and their corresponding amplitude range information in the fuzzy PID parameter increment set, perform data structure encapsulation operation, and construct a fuzzy PID parameter candidate set to be pruned that meets the input format requirements of the downstream safety pruning module, so as to provide a complete decision basis for subsequent parameter safety screening based on energy efficiency bandwidth constraints.

[0135] For the fuzzy PID parameter increment set with directional constraints and amplitude boundaries generated by S5.4, the values ​​of the proportional coefficient increment, integral coefficient increment, and derivative coefficient increment, as well as the corresponding amplitude range data, are read, and a one-to-one correspondence table between parameter increments and amplitude boundaries is established to ensure the accuracy of subsequent encapsulation.

[0136] The structured encapsulation module is invoked to insert the increments of the three types of parameters (proportion, integral, and derivative) and their respective amplitude ranges into a unified data object in the form of key-value pairs, forming a preliminary encapsulation body with field labels, data type definitions, and unit annotations.

[0137] The application parameter channel sorting rules are used to sort the fields in the initial encapsulation body according to the channel order of ratio, integral, and derivative. Channel identifiers and amplitude range labels are added to the metadata of each field to enable the downstream security trimming module to quickly index them.

[0138] Perform data consistency checks by logically comparing the upper and lower limits of the amplitude ranges for the proportional, integral, and derivative channels to ensure that the boundary values ​​of the amplitude ranges meet preset mathematical relationship conditions, such as the upper limit being greater than the lower limit and both being valid values ​​within the real number range.

[0139] After completing the consistency check, the package is converted into a serialized parameter candidate set that meets the input format requirements of the downstream safety trimming module, and a timestamp, operating condition number and graph node reference index are added to construct the fuzzy PID parameter candidate set to be trimmed.

[0140] By using serialization encapsulation and metadata enhancement processing, the incremental parameter results from the previous step are transformed into structured, traceable, and highly adaptable input data, achieving the expected technical effect of rapid identification and high-precision screening in the downstream security trimming process.

[0141] For example, in a collaborative control scenario of a central air conditioning system in an office building, the proportional coefficient increment is... The amplitude range is The increment of the integral coefficient is The amplitude range is The increment of the differential coefficient is The amplitude range is Establish channel correspondences for these three sets of data. For example, "Channel P" corresponds to the proportional coefficient increment and range, "Channel I" corresponds to the integral coefficient increment and range, and "Channel D" corresponds to the derivative coefficient increment and range. During encapsulation, add field labels such as "channel_id", "delta_value", "range_min", and "range_max", along with the unit "dimensionless coefficient" and the operating condition number "Mode_2024_01". In consistency verification, the maximum value of the amplitude range is detected. Greater than the minimum value And the maximum value of the integral coefficient increment Greater than the minimum value The parameters meet the preset boundary conditions. After serialization, a set of candidate parameters is obtained that can be directly used by the secure cropping module to perform multi-dimensional spatial mapping verification. The verification results show that the recognition time of this encapsulated data in downstream cropping processing is significantly shortened, and the screening accuracy is greatly improved.

[0142] Step S6: Based on the energy efficiency bandwidth constraints of the current chilled water system, perform safe pruning on the candidate set of fuzzy PID parameters to be pruned, eliminating parameter combinations that exceed the physical execution limit, and generating the final fuzzy PID control parameter set that meets the system stability requirements. Specifically, this includes: S6.1: Obtain the candidate set of fuzzy PID parameters to be pruned and the real-time collected operating status data of the chilled water system. Based on physical limit indicators such as the rated power of the chiller unit, the maximum frequency of the pump motor and the pressure threshold of the pipeline, calculate and generate a dynamic energy efficiency bandwidth constraint boundary set containing the upper limit of the proportional coefficient, the lower limit of the integral coefficient and the threshold of the rate of change of the derivative coefficient.

[0143] S6.2: Using the dynamic energy efficiency bandwidth constraint boundary set, perform a multi-dimensional space mapping verification operation on each set of proportional coefficient increments, integral coefficient increments, and derivative coefficient increments in the fuzzy PID parameter candidate set to be pruned, identify out-of-bounds parameter combinations located outside the dynamic energy efficiency bandwidth constraint boundary, and mark them as illegal control parameter subsets.

[0144] Using the upper limit of the proportional coefficient, the lower limit of the integral coefficient, and the threshold of the rate of change of the derivative coefficient contained in the dynamic energy efficiency bandwidth constraint boundary set, a multi-dimensional spatial mapping verification operation is performed on the candidate set of fuzzy PID parameters to be pruned to determine the legality of each parameter combination within the physical execution capability range. The increments of the proportional coefficient, integral coefficient, and derivative coefficient in the candidate set of parameters to be pruned are mapped to the corresponding constraint boundary dimensions, and a coordinate transformation matrix is ​​used to convert each increment into a normalized relative position vector for boundary judgment under a unified scale. Interval comparison operations are performed in each dimension to determine whether the increment exceeds the upper limit or falls below the lower limit, and the instantaneous change amplitude is calculated based on the rate of change threshold of the derivative coefficient to determine whether it exceeds the dynamic pressure threshold. For each parameter combination, a comprehensive three-dimensional boundary judgment vector is used to generate an overall legality label through a multi-dimensional decision function; if any dimension exceeds the boundary, the combination is marked as an illegal control parameter. The boundary judgment process is defined by the following mathematical formula: Represents the combination of parameter increments, where This is the increment of the proportional coefficient. For the increment of the integral coefficient, This represents the increment of the differential coefficient.

[0145] Define the out-of-bounds condition for the scaling factor, where This represents the upper limit of the proportionality coefficient. The boundary conditions for the integral and differential coefficients are calculated using isomorphic formulas, and are... and The system performs independent judgment. It combines the out-of-bounds Boolean values ​​from the three dimensions using a logical OR operation to generate an identifier for the subset of illegal control parameters. Through a spatial mapping verification operation, the candidate set of parameters to be pruned from the previous step is transformed into a dataset with valid labels, achieving accurate identification and classification of out-of-bounds combinations.

[0146] For example, in a commercial building with significant fluctuations in total cooling demand, a candidate set of fuzzy PID parameters to be tailored is obtained, with the increment range of the proportional coefficient being... The integral coefficient increment ranges from 0.5 to 0.8. The increment range of the differential coefficient is from 0.01 to 0.03. 0.05 to 0.07. The upper limit of the proportional coefficient obtained from real-time acquisition of chilled water system operating status data is 0.75, and the lower limit of the integral coefficient is... 0.005, the threshold for the rate of change of the differential coefficient is 0.06. The candidate set of parameters to be pruned is mapped one by one to the corresponding constraint boundary dimension. For example, a scaling factor increment of 0.8 is mapped to an upper limit of 0.75, resulting in a relative scale of... Approximately 1.067, increment of integral coefficient 0.01 is below the lower limit The increment of the differential coefficient is 0.07, which is higher than the threshold of 0.06. Based on the comprehensive logic judgment function results, this parameter combination is out of bounds and is marked as illegal. For the combination with a proportional coefficient increment of 0.6, an integral coefficient increment of 0.01, and a differential coefficient increment of 0.05, after normalizing the proportional coefficient position to 0.8, the integral coefficient position to be higher than the lower limit, and the differential coefficient position to be lower than the threshold, all three-dimensional judgments are valid, and this combination is retained in the set of valid parameters. After performing this step, the subset of illegal control parameters accounts for only about 15% of the candidate set, effectively avoiding system instability caused by parameters exceeding the physical execution limit, and ensuring that subsequent safe pruning processing can be accurately optimized within the effective parameter space.

[0147] S6.3: For the out-of-bounds parameter combinations in the illegal control parameter subset, the projection gradient descent algorithm is used to perform numerical truncation and direction correction processing, forcibly pulling the proportional coefficient increment, integral coefficient increment and differential coefficient increment that exceed the dynamic energy efficiency bandwidth constraint boundary back to the nearest legal boundary point, generating an intermediate fuzzy PID parameter correction set after preliminary correction.

[0148] For out-of-bounds parameter combinations in the illegal control parameter subset, read the numerical components of their proportional coefficient increment, integral coefficient increment, and differential coefficient increment, and calculate the difference with the corresponding legal interval in the dynamic energy efficiency bandwidth constraint boundary to establish an out-of-bounds distance vector for subsequent correction magnitude determination.

[0149] The out-of-bounds distance vector is input into the iterative framework of the projection gradient descent algorithm. The gradient projection value is calculated based on the current gradient direction and parameter space constraints, thereby realizing the directional constraint on each coefficient increment and ensuring that the algorithm always converges along the legal region of the boundary during the iteration process.

[0150] A numerical truncation rule is adopted. When the gradient projection value exceeds the boundary of the legal interval, the current incremental value is truncated according to the boundary threshold to form an intermediate correction value with boundary alignment, so as to avoid the accumulation of out-of-bounds errors caused by continued iteration.

[0151] The intermediate correction value after truncation is combined with the projection gradient direction determination information to perform direction correction processing, and the components that do not conform to the boundary orientation are adjusted to the direction that coincides with the legal region, so that the corrected parameters are consistent with the stability of the chilled water system in a physical sense.

[0152] The projection gradient update is calculated using the following mathematical expression: in, This is the current parameter increment vector. This is the step size coefficient. Given the gradient vector, the updated gradient is ensured by the projection operator. The boundary constraints are satisfied.

[0153] The parameter set that has undergone numerical truncation and direction correction is summarized to generate an intermediate fuzzy PID parameter correction set containing correction values ​​for proportional coefficient, integral coefficient, and derivative coefficient, which serves as the input for subsequent smoothing filtering.

[0154] By using projection gradient descent and truncation direction correction, the illegal control parameter combinations identified in the previous step are transformed into intermediate corrected data with valid boundaries, thereby achieving safe parameter regression and ensuring system operational stability.

[0155] For example, for a certain chilled water system, the increment of the scaling factor to be adjusted is... The upper limit of the legal range is The increment of the integral coefficient is The lower limit of the legal range is The rate of change of the differential coefficient increment is The legal threshold is The out-of-bounds distance vectors are respectively the scaling factors. Integral coefficient Differential coefficients Step size coefficient At that time, the projection gradient descent iterative calculation is as follows: the scaling factor is updated to... The integral coefficient is updated to The differential coefficients are updated to All values ​​were at valid boundary points. After the modified set was input to the smoothing filter, the chilled water system did not exhibit overshoot during sudden load changes, the pump frequency regulation process remained stable, the energy efficiency index was significantly improved compared to before, and the control performance remained stable under complex load fluctuations.

[0156] S6.4: Based on the intermediate fuzzy PID parameter correction set, combined with the thermal inertia delay characteristics and flow regulation dead zone range of the chilled water system, perform control command smoothing filtering to eliminate the risk of step response caused by parameter mutation, and output a pre-verified fuzzy PID control parameter sequence with continuity and stability.

[0157] S6.5: Perform final integrity encapsulation processing on the pre-verified fuzzy PID control parameter sequence, remove all residual abnormal values ​​and add system stability verification tags to generate the final fuzzy PID control parameter set that fully meets the current chilled water system energy efficiency bandwidth constraints and can be directly input into the fuzzy inference engine.

[0158] When encapsulating the pre-validated fuzzy PID control parameter sequence for integrity, a structured verification mechanism is used to check the validity of each proportional, integral, and derivative coefficient in the sequence, eliminating abnormal values ​​with NaN or Inf identifiers in the floating-point representation. A cross-field consistency verification function is called to perform logical association verification on the combination of proportional and integral coefficients, and the combination of proportional and derivative coefficients, ensuring the physical feasibility of parameter relationships between control channels. A numerical stability detection algorithm is introduced to calculate the gradient of parameter fluctuation amplitudes in adjacent sampling periods within the sequence. If the gradient exceeds a preset safety threshold, the corresponding parameter is marked as abnormal and filtered out. Based on the chilled water system energy efficiency bandwidth constraint library, system stability verification tags are added to the sequence. The tag generation process includes calculating the allowable parameter change rate threshold based on thermal inertia delay characteristics and hydraulic response model, and using this as the basis for tag rating. The data encapsulation module is called to integrate the verified proportional, integral, and derivative coefficients and their stability tags into a unified control parameter data structure format, outputting the final fuzzy PID control parameter set that fully conforms to the current chilled water system energy efficiency bandwidth constraints. By performing numerical validity checks, cross-field consistency verification, fluctuation gradient analysis, energy efficiency constraint label generation, and unified encapsulation processing, the results of the previous step are transformed into control parameter data with stability guarantees that can be directly input into the fuzzy inference engine, thus achieving the security and adaptability of the final output.

[0159] For example, in the chilled water system of a large commercial complex, the pre-validation fuzzy PID control parameter sequence includes a proportional coefficient range of [1.8, 2.4], an integral coefficient range of [0.15, 0.20], and a derivative coefficient range of [0.02, 0.03], with a sampling period of 2 seconds. During the numerical validity check, the combination of the proportional coefficient 2.4 and the integral coefficient 0.15 was found to satisfy the physical constraints, but the gradient of the derivative coefficient 0.03 over two consecutive periods was 0.008, which is lower than the gradient threshold of 0.01. Therefore, the stability level was marked as high. In the cross-field consistency verification, the product relationship detection formula was applied to the proportional and integral coefficients: ,in This is the proportionality coefficient. The integral coefficients are calculated as follows: The label is considered valid if it falls within the preset range [0.30, 0.40]. The allowable rate of change threshold formula is used in the generation of energy efficiency bandwidth constraint labels: ≤ The test results are The constraints are met. The final encapsulation module packages the proportional coefficient 2.2, integral coefficient 0.18, derivative coefficient 0.025, and stability label "L3" into a JSON-formatted control parameter group. After being directly input into the fuzzy inference engine, the system operates stably with rapid cooling capacity adjustment response, and the energy efficiency ratio is significantly improved.

[0160] S7: The final fuzzy PID control parameter set is input into the fuzzy inference engine, replacing the traditional fixed rule base based on error and deviation change rate. Based on the hierarchical attributes of nodes in the initial load semantic graph, hierarchical fuzzy rules are automatically generated to execute the output operations of water supply temperature setpoint adjustment and pump frequency regulation commands. Specifically, this includes: S7.1: Based on the proportional coefficient, integral coefficient and derivative coefficient values ​​in the final fuzzy PID control parameter set, and combined with the dominant semantic node and its neighborhood activation mode identified at the current moment, extract the thermal inertia level attribute, response urgency attribute and regional coupling strength attribute of the corresponding node in the load semantic map, and generate a rule construction input vector set containing multi-dimensional context features.

[0161] Based on the proportional, integral, and derivative coefficients in the final fuzzy PID control parameter set, structured data of the dominant semantic node and its neighborhood activation patterns identified at the current moment are loaded. The coefficients of the proportional, integral, and derivative channels are used as numerical features and simultaneously input into the feature fusion module along with the dominant semantic node identifier and the neighborhood node set. The node attribute records corresponding to the dominant semantic node in the load semantic graph are retrieved through a graph index, and the thermal inertia level, response urgency, and regional coupling strength attribute values ​​are extracted. These attributes are then matched with the coefficient values ​​to form a pair. A multi-dimensional feature concatenation algorithm is used to sort and concatenate the proportional, integral, and derivative coefficients with the thermal inertia level, response urgency, and regional coupling strength to generate a multi-dimensional context feature matrix. Normalization is used to perform maximum and minimum scaling on attribute values ​​of different dimensions to ensure that the thermal inertia level and coefficient values ​​are within the same numerical range, thus eliminating the interference of feature scale differences on the rule construction input vector set. Principal component analysis is performed on the context feature matrix to reduce dimensionality, retaining feature components with high correlation to the control output, forming a compact and high-information-density rule construction input vector set. Through the above processing method, the result of the previous step is transformed into a multi-dimensional feature input containing ratio, integral, differential coefficient and node context attributes, so as to realize the precise feature-driven generation of dynamic hierarchical fuzzy rules.

[0162] For example, for the central air conditioning system of a large commercial building, the final fuzzy PID control parameter set includes a proportional coefficient. Integral coefficient Differential coefficients The dominant semantic node is identified as "rapid cooling requirement," and the neighborhood activation mode contains two nodes, each with a thermal inertia level. and The response urgency attribute values ​​are respectively and The regional coupling strength attribute values ​​are respectively and During vector construction, coefficient values ​​are concatenated with node attribute values ​​to form a feature matrix. Corresponding to the dominant node, and [ Corresponding neighbor nodes. After normalization, all feature values ​​are mapped to... Within the interval, principal component analysis is used to retain the first two principal components. The resulting rule-constructed input vector set significantly improves the generation accuracy of the hierarchical fuzzy rule base. In operational verification, the hierarchical rules generated by this vector set can accurately match the thermal response requirements of different areas in a rapid cooling scenario, significantly improving the execution effect of water supply temperature setpoint adjustment and pump frequency regulation commands, shortening system response time, and improving operational stability.

[0163] S7.2: Construct an input vector set using the rules, execute a hierarchical fuzzy rule automatic generation algorithm, divide the rule level granularity according to the thermal inertia level attribute, determine the rule trigger threshold according to the response urgency attribute, and map the regional coupling strength attribute to the rule weight factor to construct a dynamic hierarchical fuzzy rule library adapted to the current operating conditions.

[0164] Based on the rule-based construction of thermal inertia level attribute data in the input vector set, the rule hierarchy partitioning module is called to perform thermal inertia grading calculation, the continuous thermal inertia level values ​​are discretized into preset hierarchical intervals, and each interval is mapped to an independent rule hierarchy identifier to establish the correspondence between hierarchy granularity and thermal response speed.

[0165] The response urgency attribute data from the rule construction input vector set is imported into the trigger threshold calculation unit. The interval partitioning function is used to generate a set of trigger thresholds corresponding to different rule levels, ensuring that high-urgency scenarios are triggered first by high-weight rules while low-urgency scenarios are delayed in triggering, thereby achieving time priority control.

[0166] The regional coupling strength attribute data in the rule construction input vector set is input into the weight mapping unit. The normalization function is called to standardize the coupling strength value to the [0,1] interval. The rule weight factor is calculated based on the nonlinear mapping relationship between coupling strength and system energy efficiency contribution. Higher execution weights are given to high coupling regions to enhance the impact of local response on global stability.

[0167] The aforementioned thermal inertia level identifiers, response urgency trigger threshold sets, and regional coupling strength weight factors are integrated into a rule generation parameter matrix. Through a dynamic hierarchical fuzzy rule automatic generation algorithm, a corresponding fuzzy logic condition set is constructed for each level of granularity, and weight factors are embedded in the rule entries as adjustment parameters for inference priority, thereby realizing the ability of the rule base to dynamically match the current operating conditions.

[0168] The rule generation algorithm outputs a dynamic hierarchical fuzzy rule base containing rule entries, which transforms the multidimensional context features of the previous step into a rule set with hierarchical granularity, triggering conditions and weight factors, so as to achieve accurate adaptation to the operating conditions.

[0169] For example, in the central air conditioning system of a large commercial building, the thermal inertia level attribute of the rule construction input vector set ranges from 2.5 to 5.0. The rule hierarchy partitioning module discretizes it into three levels: low thermal inertia (2.5~3.0), medium thermal inertia (3.1~4.0), and high thermal inertia (4.1~5.0). The response urgency attribute ranges from 0 to 10. The trigger threshold calculation unit sets the high urgency threshold to 8, the medium urgency threshold to 5, and the low urgency threshold to 2. The regional coupling strength attribute ranges from 0.3 to 0.9. After normalization, it is mapped to the weight factor interval [0.33, 1.0], based on the nonlinear mapping function. Calculation, where The normalization coefficient is... The weight calculation result for the high-coupling region is 0.81, while the weight calculation result for the low-coupling region is only 0.11. The dynamic hierarchical fuzzy rule automatic generation algorithm generates weak integral-strong differential strategy rule entries in the high thermal inertia level and assigns a weight factor of 0.81 to the high-coupling region for priority execution; in the low thermal inertia level, it generates strong proportional-weak differential strategy rule entries and assigns a weight factor of 0.33 to reduce execution priority. The final output dynamic hierarchical fuzzy rule library successfully matched 85% of the real-time operating conditions during the testing period, significantly improving the system's response coordination in high thermal inertia multi-coupling scenarios and maintaining stable operation in low thermal inertia low-coupling scenarios.

[0170] S7.3: Based on the dynamic hierarchical fuzzy rule library, the real-time collected chilled water supply and return water temperature difference signal and flow deviation signal are fuzzified to obtain a temperature difference fuzzy subset and a flow fuzzy subset. The matching rule entries in the dynamic hierarchical fuzzy rule library are called to perform logical reasoning operations to generate a fuzzy control quantity output set that reflects the system adjustment trend.

[0171] The real-time collected chilled water supply and return temperature difference signals and flow deviation signals are used as input data, corresponding to the current deviation status of the temperature control channel and flow control channel, respectively. The supply and return water temperature difference signals are fuzzified, and based on a preset temperature difference membership function, continuous values ​​are mapped to fuzzy subsets of temperature differences. The membership function parameters are determined according to the building's thermal inertia level, and membership values ​​for low, medium, and high temperature differences are calculated for different thermal inertia regions. Similarly, the flow deviation signals are fuzzified, and based on a preset flow membership function, continuous values ​​are mapped to fuzzy subsets of flow. The shape parameters of the membership function are determined according to the response urgency attribute, and membership values ​​for low, medium, and high deviations are calculated for different urgency levels. The temperature difference fuzzy subset and the flow fuzzy subset are combined to form input fuzzy variable pairs, which serve as trigger nodes for dynamic hierarchical fuzzy rule base matching operations. The system calls upon rule entries from a dynamic hierarchical fuzzy rule base that match the current fuzzy variable pair. Logical reasoning is performed based on rule weight factors. During reasoning, the maximum membership method is used to activate and determine the output fuzzy quantities. Furthermore, the system utilizes rule hierarchy granularity to perform hierarchical synthesis of outputs at different thermal inertia levels, generating a set of fuzzy control outputs reflecting the system's adjustment trend. Through a chain-like processing of fuzzification and rule matching, real-time physical deviations are transformed into context-adaptive fuzzy control trend signals, enabling accurate determination of the direction of water supply temperature and pump frequency adjustment.

[0172] For example, in a large commercial building, the real-time temperature difference between the chilled water supply and return is 1.8℃, the flow rate deviation is 0.25 m³ / h, the thermal inertia level is set to high, and the response urgency level is set to medium. The temperature difference membership function uses a trigonometric function, with the peak value for the low temperature difference at 0.5℃, the peak value for the medium temperature difference at 1.5℃, and the peak value for the high temperature difference at 3.0℃. The membership value calculation formula is as follows: in, Input value for temperature difference. Center of the peak, This is the half-width parameter. Substitute it into the peak center of the high temperature difference. ℃ and half-width parameters Calculated in ℃, the membership value of the temperature difference is The membership function for flow rate is a Gaussian function, and the formula is as follows: in, This is the flow deviation value. Center of the peak, Let the standard deviation parameter be substituted. Substitute the peak center of the high deviation parameter. m³ / h and standard deviation Calculation of m³ / h, flow membership value The temperature difference fuzzy subset and the flow rate fuzzy subset are combined, and the rule "medium temperature difference and medium deviation → slight increase in temperature and slight increase in pump frequency" is matched through a dynamic hierarchical fuzzy rule base. Maximum membership degree synthesis is then performed, and the highest membership degree value of the temperature control quantity in the output fuzzy control quantity set is [value missing]. The pump frequency control variable has a high membership value. The verification results show that the fuzzy control trend signal output by this step can significantly improve the system's response efficiency to medium temperature difference and medium deviation after the equipment is executed, and maintain a stable match between cooling supply and terminal demand.

[0173] S7.4: Perform defuzzification processing on the fuzzy control output set, and perform weighted correction on the parameter increment direction and amplitude range in the final fuzzy PID control parameter group to calculate the water supply temperature setpoint adjustment and chilled pump frequency adjustment with clear physical meaning, forming a preliminary equipment execution instruction sequence.

[0174] S7.5: Based on the preliminary equipment execution instruction sequence, perform instruction smoothing filtering and safety limiting processing to eliminate high-frequency jitter components and ensure that the adjustment range is within the safe operating bandwidth of the chiller and pump group, generate the final water supply temperature setpoint adjustment instruction and pump group frequency adjustment instruction, and complete the coordinated control output of the source side equipment.

[0175] The system receives the initial equipment execution command sequence as input data, analyzes the numerical components of the water supply temperature setpoint adjustment and chilled water pump frequency adjustment contained therein, and establishes a unified command data structure for signal processing. The adjustment signal undergoes smoothing processing based on a finite impulse response (FIR) filter, with filter coefficients preset according to the thermal inertia delay characteristics of the chilled water system to reduce jitter caused by high-frequency components. For the smoothed signal, a Fast Fourier Transform (FFT) analysis module is invoked to identify its spectral components, locate components exceeding a preset high-frequency threshold, and implement a phase cancellation strategy to further suppress residual oscillations. The smoothed signal is read and its amplitude is compared with the safe operating bandwidth data of the chiller and pump units. If the amplitude exceeds the allowable range, amplitude limiting is performed. The amplitude limiting process uses a projection function to map the amplitude value within the bandwidth boundary, ensuring that parameter adjustments do not disrupt system stability. After amplitude limiting, the time series of the adjustment signal is reconstructed, and time delay compensation calculations are performed in conjunction with the equipment execution cycle to ensure that the output command matches the response rhythm of the source-side equipment. Through the aforementioned chain-like processing, the initial equipment execution command sequence is transformed into a final water supply temperature setpoint adjustment command and a pump frequency adjustment command, possessing smoothness, amplitude safety, and responsiveness, thereby achieving the goal of coordinated control of the source-side equipment. For example, in the central air conditioning system of a large commercial building, the initial equipment execution command sequence includes a water supply temperature adjustment of 0.8℃ and a chilled pump frequency adjustment of 12Hz. The filtering stage employs an FIR filter of order 5, with coefficients set to [0.1, 0.15, 0.4, 0.15, 0.1] based on the building's thermal inertia characteristics. After filtering, the spectrum analysis stage detects frequency components higher than 5Hz in the adjustment commands, and phase reversal is implemented for this high-frequency band to eliminate jitter. The safe operating bandwidth in the amplitude limiting stage is set to a water supply temperature adjustment not exceeding ±1℃ and a chilled pump frequency adjustment not exceeding ±15Hz. After amplitude projection mapping, the adjustment amounts are corrected to 0.8℃ and 12Hz respectively, without triggering amplitude reduction. The time delay compensation stage performs compensation based on a pump unit response delay of 0.2s, adjusting the command time sequence to the right accordingly to synchronize the control action with the equipment response. In the technical processing, the limiting algorithm can be expressed as: in, This is the current adjustment amount. This is the lower limit of bandwidth. This is the upper limit of the bandwidth; this process ensures that the output falls within the safe operating range of the equipment. After executing the above chain derivation, the final generated water supply temperature setpoint adjustment command is 0.8℃ and the pump group frequency adjustment command is 12Hz. In actual operation, the system exhibits stable response and no high-frequency oscillation, and the source-load coordinated control effect is significantly improved.

[0176] S8: Monitor the jump status of the dominant semantic node and the distribution deviation of the neighborhood activation mode within a continuous sampling period. If the distribution deviation is detected to meet a preset threshold condition, trigger the online fine-tuning of the initial load semantic map and the recalibration process of the nonlinear mapping network to complete the adaptive closed-loop optimization of the control strategy. Specifically, this includes: S8.1: Obtain the dominant semantic node sequence and neighborhood activation mode distribution data within a continuous sampling period, perform temporal difference calculation on the dominant semantic node sequence to generate node jump feature vectors, and perform probability density estimation on the neighborhood activation mode distribution data to generate the neighborhood distribution fingerprint at the current moment, thereby completing the basic feature extraction of the dynamic changes in semantic state.

[0177] S8.2: Based on the node jump feature vector, execute the boundary detection algorithm to identify semantic mutation events under abnormal operating conditions. At the same time, use the Kalman divergence formula to calculate the statistical distance between the neighborhood distribution fingerprint and the historical baseline distribution to generate a semantic distribution deviation index, thereby quantifying the drift of the current operating condition relative to the historical normal.

[0178] S8.3: Based on the semantic distribution deviation index and the preset dynamic threshold conditions, a logical comparison is performed. If the semantic drift triggering condition is met, an online fine-tuning instruction for the graph is generated. Otherwise, the current initial load semantic graph and nonlinear mapping network operation status are maintained to determine whether to start the adaptive closed-loop optimization process.

[0179] S8.4: In response to the online fine-tuning instruction of the graph, the newly collected abnormal semantic fragments are fused into the initial load semantic graph using an incremental graph embedding algorithm to update the node attributes and edge weights. Based on the updated topology, a backpropagation recalibration operation is performed on the nonlinear mapping network to correct the parameter mapping function, thereby completing the online iterative upgrade of the control strategy model.

[0180] In one embodiment, a central air conditioning source-load coordinated control system is provided, comprising: a terminal energy consumption characteristic sequence acquisition module, a basic semantic tag set acquisition module, an initial load semantic map construction module, a load semantic representation vector acquisition module, a fuzzy PID parameter candidate set acquisition module, a final fuzzy PID control parameter group acquisition module, and a regulation command output module, wherein: Terminal energy consumption feature sequence acquisition module: used to acquire the raw data stream collected by multi-source sensors at the terminal of central air conditioning, and to perform spatiotemporal alignment and cleaning processing on the raw data stream to generate terminal energy consumption feature sequence; Basic semantic tag set acquisition module: Based on the end-point energy consumption feature sequence, combined with building functional zoning information and personnel activity heat map distribution, semantic annotation processing is performed on the time-series segments in the original data stream to define basic semantic units and generate a basic semantic tag set; Initial load semantic graph construction module: used to construct an initial load semantic graph by utilizing the causal relationships between the basic semantic tag set and historical operation data; Load semantic representation vector acquisition module: used to embed the end-point energy consumption feature sequence into the initial load semantic map, and to identify the dominant semantic nodes and their neighborhood activation modes under the current operating conditions through a graph neural network model to obtain the load semantic representation vector; The module for obtaining the candidate set of fuzzy PID parameters to be pruned is used to perform dynamic mapping operation of fuzzy PID control parameters based on the load semantic representation vector and the pre-trained nonlinear mapping network. According to the thermal inertia level and response urgency in the node attributes of the initial load semantic map, the module calculates the direction and amplitude range of the proportional coefficient increment, integral coefficient increment and differential coefficient increment, and generates the candidate set of fuzzy PID parameters to be pruned. The final fuzzy PID control parameter set acquisition module is used to perform pruning processing on the fuzzy PID parameter candidate set to be pruned according to the energy efficiency bandwidth constraints of the current chilled water system, and generate the final fuzzy PID control parameter set. Adjustment command output module: used to input the final fuzzy PID control parameter group to the fuzzy inference engine, automatically generate hierarchical fuzzy rules based on the hierarchical attributes of the nodes in the initial load semantic graph, and execute the output operation of water supply temperature setpoint adjustment and pump group frequency adjustment commands.

[0181] This application provides a central air conditioning source-load coordinated control method, which can be applied to, for example... Figure 4 In the application environment shown, terminal 102 communicates with server 104 via a network. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices, and server 104 can be a standalone server or a server cluster consisting of multiple servers.

[0182] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data generated during the implementation of a central air conditioning source-load coordinated control method. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a central air conditioning source-load coordinated control method.

[0183] In one embodiment, a computer device is provided, the computer device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of any of the methods in the above method embodiments.

[0184] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of any of the methods described in the above method embodiments.

[0185] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0186] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0187] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0188] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for coordinated control of source and load in a central air conditioning system, characterized in that, Specifically, it includes: S1: Acquire the raw data stream collected by the multi-source sensors at the central air conditioning terminal, and perform spatiotemporal alignment and cleaning processing on the raw data stream to generate a terminal energy consumption characteristic sequence; S2: Based on the end-point energy consumption feature sequence, combined with building functional zoning information and personnel activity heat map distribution, perform semantic annotation processing on the time-series segments in the original data stream and define basic semantic units to generate a basic semantic tag set; S3: Construct an initial load semantic map by utilizing the causal relationships between the basic semantic tag set and historical operational data; S4: Embed the end-point energy consumption feature sequence into the initial load semantic map, and use a graph neural network model to identify the dominant semantic nodes and their neighborhood activation modes under the current operating conditions to obtain the load semantic representation vector; S5: Based on the load semantic representation vector and the pre-trained nonlinear mapping network, perform dynamic mapping operation on the fuzzy PID control parameters to generate a candidate set of fuzzy PID parameters to be pruned. S6: Based on the energy efficiency bandwidth constraints of the current chilled water system, the fuzzy PID parameter candidate set to be trimmed is trimmed to generate the final fuzzy PID control parameter set. S7: Input the final fuzzy PID control parameter set into the fuzzy inference engine, automatically generate hierarchical fuzzy rules based on the hierarchical attributes of the nodes in the initial load semantic graph, and execute the output operations of water supply temperature setpoint adjustment and pump group frequency adjustment commands.

2. The central air conditioning source-load coordinated control method according to claim 1, characterized in that, Following S7, S8 is also included: monitoring the jump status of the dominant semantic node and the distribution deviation of the neighborhood activation mode within a continuous sampling period; if the distribution deviation meets a preset threshold condition, the online fine-tuning of the initial load semantic map and the recalibration process of the nonlinear mapping network are triggered.

3. The central air conditioning source-load coordinated control method according to claim 1, characterized in that, S3 specifically includes: Obtain the basic semantic tag set and historical running data, and calculate the co-occurrence frequency of any two basic semantic units in the historical running data within the same time window to generate a basic semantic co-occurrence frequency matrix; Based on the co-occurrence frequency matrix of the basic semantics, the successive relationship between the basic semantic units is quantified and calculated to determine the conditional probability value of the transition from the previous basic semantic unit to the next basic semantic unit, and a set of basic semantic temporal transition probabilities is generated. Based on the set of basic semantic temporal transition probabilities and the temperature difference change rate and valve action delay data in historical operation data, the driving and driven relationships between the basic semantic units are significantly verified, and a list of basic semantic causal influence strengths is generated. Using the basic semantic units as a vertex set, and combining the basic semantic co-occurrence frequency matrix, the basic semantic temporal transition probability set, and the basic semantic causal influence intensity list as edge attribute data, a topology assembly operation of a directed weighted graph is performed to generate the initial load semantic graph topology structure. Based on the initial load semantic graph topology, metadata enhancement processing is performed on each graph node to generate the final initial load semantic graph containing complete node attributes and edge weights.

4. The central air conditioning source-load coordinated control method according to claim 3, characterized in that, S4 specifically includes: The end-point energy consumption feature sequence is obtained, and node matching calculation is performed on the time segment in the end-point energy consumption feature sequence based on the definition rules of the basic semantic tag set to obtain the initial semantic node embedding vector. Using the initial semantic node embedding vector as input, the graph neural network model is invoked to perform message passing and aggregation operations. The hidden state of the nodes is updated according to the edge weights defined in the initial load semantic graph, and a dynamic node feature representation that integrates temporal transition probability and causal influence intensity is generated. Based on the dynamic node feature representation, an attention mechanism is performed to calculate the score, and the node with the highest activation score is selected as the dominant semantic node. The set of neighboring nodes within a preset hop count range of the dominant semantic node is extracted to construct a local subgraph topology. The set of neighboring nodes in the local subgraph topology is subjected to feature weighted fusion processing. Combining the attribute information of the dominant semantic node with the spatial distribution characteristics of the neighborhood activation mode, the high-dimensional topological information is compressed, and the load semantic representation vector is output.

5. The central air conditioning source-load coordinated control method and system according to claim 1, characterized in that, S5 specifically includes: The load semantic representation vector and the graph node attribute data containing thermal inertia level and response urgency attributes are obtained. A nonlinear mapping network model is constructed using a multilayer perceptron architecture. The load semantic representation vector is then subjected to high-dimensional feature space projection processing to generate a hidden layer activation state matrix. Based on the hidden layer activation state matrix and the pre-trained weight coefficient matrix, a weighted summation operation is performed on the fully connected layer to output the original parameter offset sequence containing the proportional channel, integral channel and differential channel. The original parameter offset sequence is read and combined with the thermal inertia level. The original parameter offset sequence is then subjected to amplitude correction processing to generate an intermediate parameter adjustment vector with thermal inertia adaptability. Based on the intermediate parameter adjustment vector and the response urgency attribute, the direction determination logic function is used to calculate the change trend of each control channel, determine the incremental direction of the proportional coefficient, the incremental direction of the integral coefficient and the incremental direction of the derivative coefficient, and combine the preset step size limit threshold to calculate the corresponding amplitude range, and generate a fuzzy PID parameter increment set. By integrating the incremental directions of the proportional coefficient, integral coefficient, and derivative coefficient, as well as their corresponding amplitude ranges, from the incremental set of fuzzy PID parameters, a data structure encapsulation operation is performed to construct a candidate set of fuzzy PID parameters to be pruned.

6. The central air conditioning source-load coordinated control method according to claim 1, characterized in that, The raw data stream includes: the difference between the indoor temperature setpoint and the actual temperature, the start / stop status of the fan coil unit, the water valve opening degree, and timestamp information.

7. The central air conditioning source-load coordinated control method according to claim 1, characterized in that, The basic semantic units include: rapid cooling requirements, steady-state maintenance requirements, local overcooling suppression, and low-load hibernation at night.

8. A central air conditioning source-load coordinated control system, characterized in that, include: Terminal energy consumption feature sequence acquisition module: used to acquire the raw data stream collected by multi-source sensors at the terminal of central air conditioning, and to perform spatiotemporal alignment and cleaning processing on the raw data stream to generate terminal energy consumption feature sequence; Basic semantic tag set acquisition module: Based on the end-point energy consumption feature sequence, combined with building functional zoning information and personnel activity heat map distribution, semantic annotation processing is performed on the time-series segments in the original data stream to define basic semantic units and generate a basic semantic tag set; Initial load semantic graph construction module: used to construct an initial load semantic graph by utilizing the causal relationships between the basic semantic tag set and historical operation data; Load semantic representation vector acquisition module: used to embed the end-point energy consumption feature sequence into the initial load semantic map, and to identify the dominant semantic nodes and their neighborhood activation modes under the current operating conditions through a graph neural network model to obtain the load semantic representation vector; The module for obtaining the candidate set of fuzzy PID parameters to be pruned is used to perform dynamic mapping operation of fuzzy PID control parameters based on the load semantic representation vector and the pre-trained nonlinear mapping network. According to the thermal inertia level and response urgency in the node attributes of the initial load semantic map, the module calculates the direction and amplitude range of the proportional coefficient increment, integral coefficient increment and differential coefficient increment, and generates the candidate set of fuzzy PID parameters to be pruned. The final fuzzy PID control parameter set acquisition module is used to perform pruning processing on the fuzzy PID parameter candidate set to be pruned according to the energy efficiency bandwidth constraints of the current chilled water system, and generate the final fuzzy PID control parameter set. Adjustment command output module: used to input the final fuzzy PID control parameter group to the fuzzy inference engine, automatically generate hierarchical fuzzy rules based on the hierarchical attributes of the nodes in the initial load semantic graph, and execute the output operation of water supply temperature setpoint adjustment and pump group frequency adjustment commands.

9. A computer device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor; when executed by the processor, the computer program implements the steps of a central air conditioning source-load coordinated control method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of a central air conditioning source-load coordinated control method as described in any one of claims 1 to 7.

Images