Central air conditioning fan coil intelligent regulation and control method

Abstract

The present application relates to a kind of central air conditioning fan coil intelligent control method, it is related to Internet of Things technology, for the problems such as large noise of multi-source environmental parameter data, inaccurate working condition modeling and the non-synergistic of equipment adjustment response and energy consumption optimization, multi-parameter distributed sensing, data preprocessing, feature extraction and integrated control system of cross-domain knowledge transfer are proposed.By multidimensional real-time acquisition of room temperature and humidity, fan current, valve opening, passenger flow, external weather and building information, combined with denoising, normalization, alignment and missing completion, a high-quality structured data set is formed.Using multi-objective neural network and digital twin simulation, load trend prediction, equipment anomaly early warning and adjustment strategy virtual deduction are realized, and the best adjustment timing window and activation threshold are output, and dynamic adaptive control is realized by combining reinforcement learning.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and prediction of central air conditioning systems using Internet of Things (IoT) technology, and particularly to an intelligent control method for central air conditioning fan coil units. Background Technology

[0002] Central air conditioning systems are widely deployed in buildings and commercial office buildings, with fan coil units playing a crucial role in regulating indoor temperature and humidity and achieving regional microenvironment comfort. As industry demands for smart buildings and green energy conservation continue to rise, intelligent optimization and control of central air conditioning terminals are gradually becoming a key focus within the industry.

[0003] Some systems have also explored data aggregation and distributed control strategies based on interconnected IoT platforms, possessing basic functions such as fault diagnosis and adaptive energy-saving modes. While there are basic automatic control and some intelligent optimization applications, the representative solutions of existing systems mainly focus on local optimization in single scenarios: on the one hand, data-driven intelligent control models rely heavily on locally accumulated historical data for learning, resulting in a relatively slow parameter "cold start" phase, making it difficult to quickly achieve ideal performance in new scenarios, atypical rooms, or equipment aging conditions; on the other hand, existing predictive models have limited ability to anticipate and rapidly adjust to sudden load changes, spatial dynamics, and external environmental changes, with strategy updates lagging behind sudden changes in actual energy consumption and comfort demands. Current data-driven control strategies often exhibit a "lag" in strategy correction compared to demand changes, making it difficult to proactively predict and respond promptly to conditions such as sudden load changes, extreme weather impacts, or surges in population density; for long-term dynamics such as equipment aging and abnormal trends, the control system often reacts with lag, leading to high energy consumption, large fluctuations in comfort, and poor user experience. Summary of the Invention

[0004] This application provides a method for intelligent control of central air conditioning fan coil units, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0005] This application provides a method for intelligent control of central air conditioning fan coil units, specifically including:

[0006] S1: Collect environmental parameter data from multiple locations within the terminal of the central air conditioning fan coil unit. The environmental parameter data includes room temperature and humidity, fan current, valve opening degree, equipment surface temperature, room occupancy statistics, external meteorological data, and building orientation information.

[0007] S2: Perform time-series denoising, normalization, and outlier removal on the environmental parameter data to obtain a dataset suitable for cross-scene modeling.

[0008] S3: Based on the preprocessed dataset, a feature extraction process for multi-source information fusion is constructed for different room types, area attributes and equipment aging status. Load change features, comfort indicators and energy consumption trend features are extracted, and multi-dimensional operating condition labels are generated.

[0009] S4: Input the multi-dimensional working condition labels into the cross-domain knowledge transfer model, and use the energy consumption and comfort adjustment parameters learned in the known scenario to perform knowledge transfer and parameter pre-initialization in order to obtain the parameter priors of the current target environment.

[0010] S5: Based on a neural network time-series prediction model, combined with internal sensor data and external dynamic data, the load trend and equipment anomaly probability of the room within a future preset time window are predicted, and the load mutation risk value and anomaly tendency characteristics are output.

[0011] S6: Construct a real-time digital twin simulation unit, and perform virtual calculations on the execution consequences of various adjustment strategies based on the prior parameters and mutation risk values, predict the impact of fan coil unit adjustment actions on temperature control response and energy consumption, and determine the optimal adjustment timing window and activation threshold.

[0012] S7: Using the optimal adjustment timing window and activation threshold as input, dynamically optimize the policy network in the reinforcement learning controller, and optimize the output pre-execution control command of the fan coil unit based on the three-element adaptive adjustment mechanism of "expected energy consumption benefit - policy execution lag - comfort tolerance".

[0013] S8: In accordance with the pre-execution control instructions, the operation commands are issued to the central air conditioning fan coil unit terminals in real time to activate the heating, cooling and fan speed adjustment commands in advance, so that the system can dynamically respond to the predicted load changes or equipment abnormalities.

[0014] S9: After periodic monitoring and control, the actual feedback data is compared with the aforementioned prediction results to analyze the actual changes in energy consumption, comfort level, and dynamic occurrence of anomalies. Through self-evolutionary learning, the strategy network and transfer model parameters are optimized to achieve a closed loop of predictive intelligent control for central air conditioning fan coil units.

[0015] The intelligent control method for central air conditioning fan coil units provided in this application has the following beneficial effects:

[0016] (1) This invention utilizes a multi-point sensor network deployed at the end of the fan coil unit to combine multi-dimensional data such as room traffic, external weather, and building attributes to achieve spatiotemporal structured perception of environmental parameters in various areas of the room. Combined with multi-layer preprocessing algorithms such as temporal denoising, normalization, outlier removal, and missing value completion, it comprehensively eliminates equipment differences, sensor errors, and data anomalies, significantly improving data reliability and providing a robust foundation for subsequent modeling and intelligent decision-making. Actual test data shows that after data processing using this invention, the noise variance of environmental parameters is significantly reduced, the anomaly detection rate is less than 0.5%, and the data availability is significantly higher than traditional solutions.

[0017] (2) Through clustered and partitioned feature mapping and a multi-dimensional operating condition labeling system, this invention can automatically archive high-dimensional features for different room types, spatial attributes and equipment health status, and achieve comprehensive extraction of load changes, comfort indicators and energy consumption trends. By utilizing cross-domain adaptive transfer learning and distributed consistency parameter mapping, this invention can efficiently transfer energy-saving control experience and comfort adjustment weights from historical scenarios to new environments, and achieve rapid cold start and high-quality initial strategy pre-response.

[0018] (3) By using neural networks to jointly predict load trends, comfort levels, and equipment anomaly risks in a time series, it is possible to effectively detect load changes and the probability of abnormal events 5-30 minutes in advance. The prediction error is significantly reduced compared to existing single-objective prediction methods, and it is more sensitive to changes in macro and micro environments and sudden changes in personnel behavior.

[0019] (4) The lightweight digital twin simulation unit constructed realizes the virtual calculation of the response results of different adjustment strategies. It can automatically deduce temperature changes, energy consumption curves and time delay effects, which greatly improves the scientific nature and forward-looking nature of strategy adjustment. The adjustment timing window and activation threshold determined by the simulation results can minimize the lag in strategy execution. Attached Figure Description

[0020] Appendix Figure 1 This is the main flowchart of a central air conditioning fan coil unit intelligent control method.

[0021] Appendix Figure 2 This is a sub-flowchart of a central air conditioning fan coil unit intelligent control method.

[0022] Appendix Figure 3 This is another sub-flowchart of a central air conditioning fan coil unit intelligent control method. Detailed Implementation

[0023] 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.

[0024] 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. In addition, examples of various specific processes and materials are provided in this invention, but those skilled in the art will recognize the application of other processes and the use of other materials.

[0025] As attached Figure 1 As shown, this application provides a method for intelligent control of central air conditioning fan coil units, specifically including:

[0026] S1: Collect environmental parameter data from multiple locations within the terminal of the central air conditioning fan coil unit. The environmental parameter data includes room temperature and humidity, fan current, valve opening degree, equipment surface temperature, room occupancy statistics, external meteorological data, and building orientation information.

[0027] S2: Perform time-series denoising, normalization, and outlier removal on the environmental parameter data to obtain a dataset suitable for cross-scene modeling.

[0028] S3: Based on the preprocessed dataset, a feature extraction process for multi-source information fusion is constructed for different room types, area attributes and equipment aging status. Load change features, comfort indicators and energy consumption trend features are extracted, and multi-dimensional operating condition labels are generated.

[0029] S4: Input the multi-dimensional working condition labels into the cross-domain knowledge transfer model, and use the energy consumption and comfort adjustment parameters learned in the known scenario to perform knowledge transfer and parameter pre-initialization in order to obtain the parameter priors of the current target environment.

[0030] S5: Based on a neural network time-series prediction model, combined with internal sensor data and external dynamic data, the load trend and equipment anomaly probability of the room within a future preset time window are predicted, and the load mutation risk value and anomaly tendency characteristics are output.

[0031] S6: Construct a real-time digital twin simulation unit, and perform virtual calculations on the execution consequences of various adjustment strategies based on the prior parameters and mutation risk values, predict the impact of fan coil unit adjustment actions on temperature control response and energy consumption, and determine the optimal adjustment timing window and activation threshold.

[0032] S7: Using the optimal adjustment timing window and activation threshold as input, dynamically optimize the policy network in the reinforcement learning controller, and optimize the output pre-execution control command of the fan coil unit based on the three-element adaptive adjustment mechanism of "expected energy consumption benefit - policy execution lag - comfort tolerance".

[0033] S8: In accordance with the pre-execution control instructions, the operation commands are issued to the central air conditioning fan coil unit terminals in real time to activate the heating, cooling and fan speed adjustment commands in advance, so that the system can dynamically respond to the predicted load changes or equipment abnormalities.

[0034] S9: After periodic monitoring and control, the actual feedback data is compared with the aforementioned prediction results to analyze the actual changes in energy consumption, comfort level, and dynamic occurrence of anomalies. Through self-evolutionary learning, the strategy network and transfer model parameters are optimized to achieve a closed loop of predictive intelligent control for central air conditioning fan coil units.

[0035] Step S1: Collect environmental parameter data from multiple locations within the central air conditioning fan coil unit terminal. This environmental parameter data includes room temperature and humidity, fan current, valve opening degree, equipment surface temperature, room occupancy statistics, external meteorological data, and building orientation information. Specifically, this includes:

[0036] S1.1: Distributed data acquisition is performed on the room temperature and humidity sensor network to obtain temperature and relative humidity data at each sampling point in the room, forming a raw parameter matrix of room temperature and humidity, and realizing structured perception of the spatiotemporal distribution of environmental temperature and humidity.

[0037] The input data includes the raw analog or digital signals output by the temperature and humidity sensor network distributed throughout the room, as well as the spatial layout diagram of the central air conditioning fan coil unit terminals and the corresponding sampling point numbers.

[0038] A distributed multi-point temperature and humidity acquisition method is adopted (parameters: N≥5 detection points in a typical room layout, temperature accuracy ±0.1℃, humidity accuracy ±1.5%RH) to achieve real-time acquisition of environmental parameters at different spatial locations in the room.

[0039] Furthermore, by using bus-based communication (such as RS485 / BACnet) or wireless networking (such as ZigBee, LoRa, etc.), the data from all distributed temperature and humidity probes are synchronously transmitted to the data acquisition main control unit, and time-stamped to obtain the synchronously sampled multi-node raw temperature and humidity data stream.

[0040] A multi-channel time-series buffering and periodic polling acquisition strategy is adopted to automatically collect temperature and humidity data from each sampling point by channel and perform short-time averaging, thereby achieving local filtering and effective sampling of data redundancy.

[0041] Furthermore, using a spatial structure mapping algorithm, the temperature and humidity data from each sampling point are assigned according to their physical spatial coordinates and integrated into a structured raw parameter matrix of room temperature and humidity, as follows:

[0042]

[0043] in, Indicates the first Sampling point time The temperatures, i = 1, 2, 3, ..., N; where This represents the total number of sampling points.

[0044] Indicates the first Sampling point time The relative humidity values, i = 1, 2, 3, ..., N; where This represents the total number of sampling points.

[0045] By utilizing spatial interpolation and trend statistics algorithms for temperature and humidity data (such as Kriging interpolation or inverse distance weighted IDW), a continuous temperature and humidity distribution field can be reconstructed within the three-dimensional structural domain of a room, enabling structured perception of spatiotemporal distribution characteristics.

[0046] Through the above chain derivation, the raw temperature and humidity data collected from the distribution are transformed into a high-dimensional, structured raw temperature and humidity parameter matrix with time and space labels, which provides support for the environmental monitoring of the central air conditioning fan coil unit terminal and improves the system's perception accuracy of changes in the room environment.

[0047] For example, six temperature and humidity sampling points are set up in a standard 15 m² office room, with a sampling cycle of 30 s / time. The sensor model is set to industrial grade SHT35 (temperature resolution 0.01℃, humidity resolution 0.01%RH). The main control unit synchronously reads the values ​​of each node according to a predetermined protocol, forming the following temperature and humidity parameter matrix:

[0048]

[0049] Using the Kriging spatial interpolation method, the distributed point data is mapped to the entire room area, reconstructing a spatial variation distribution surface with a temperature gradient of 0.5℃ and a humidity spatial variation of no more than 2%. Sampling results show that the system can output a high-precision, structured temperature and humidity distribution parameter matrix within a delay of less than 30 seconds, providing complete and accurate data support for subsequent dynamic energy consumption prediction and comfort assessment.

[0050] S1.2: Perform synchronous data acquisition operation on the wind turbine current sensor acquisition module. By acquiring the current signal of the wind turbine motor in real time, obtain the raw parameter array of the wind turbine current, and realize the real-time quantification of the wind turbine's operating status and energy consumption level.

[0051] S1.3: Collect and transmit signals from the intelligent terminal for valve opening, obtain the opening data of the fan coil unit valve, form a valve opening parameter chain list, and accurately quantify the execution status of the terminal cold and hot medium flow regulation.

[0052] S1.4: Perform periodic temperature measurement and acquisition on the distributed temperature sensor array on the equipment surface to realize the time-series aggregation of the raw temperature data of the equipment surface, which is used for subsequent judgment of the equipment's heat dissipation and abnormal operation trends.

[0053] S1.5: Capture data from the room traffic statistics camera module or infrared sensor array, and obtain the time series of room traffic statistics parameters through the people detection algorithm to characterize the room load demand and usage scenario.

[0054] The input data includes the digital signal output of the pedestrian flow statistics camera module or infrared sensor array configured in typical passage areas such as the ceiling, doorway, and corridor, as well as the spatial layout diagram and sampling point number of the central air conditioning fan coil unit terminal.

[0055] A high-precision people detection algorithm is adopted (parameter settings: camera frame rate ≥15 fps, infrared array resolution 16×16, statistical cycle 30 s / time) to realize the function of detecting and counting human objects in the main passage and entrance area of ​​the room.

[0056] Furthermore, by employing convolutional neural networks (CNN) or YOLOv5 object detection algorithms, combined with region of interest (ROI) segmentation methods, the acquired camera images or infrared dot array signals are analyzed in real time to identify human targets in the images / arrays and output the number of human targets in each frame. .

[0057] Timing buffering and moving average processing are used (sliding window length) (Frames), by counting the number of human targets in consecutive frames, robust temporal parameters of pedestrian flow are obtained:

[0058]

[0059] in, For a moment The corresponding smoothed statistics of pedestrian flow. Let K be the number of human targets detected at time tk, where K is the time offset. The normalization coefficient is used to average the accumulated results, so that the output is the "average number of human bodies" within these W frames.

[0060] Furthermore, through a spatial layout mapping algorithm, the pedestrian flow statistics of each sampling point or monitoring section are archived according to the physical space division to form a time-series vector of pedestrian flow parameters for the entire room, which facilitates subsequent load demand modeling and scene recognition.

[0061] By employing identity deduplication and dwell time threshold elimination methods, duplicate or long-staying human targets are filtered out, improving the uniqueness and timeliness of people flow statistics parameters and enabling a more accurate depiction of actual room load demand.

[0062] Through the above algorithm processing, the sensor or camera signals from the previous step are transformed into structured time-series data of people flow statistics, realizing high-precision dynamic representation of room usage scenarios and real-time load requirements.

[0063] For example, in a 20 m² conference room, one infrared people flow array sensor (16×16 resolution, 30 fps refresh rate) is deployed at the entrance and one at the central aisle. Statistical results are collected every 60 seconds. Using a real-time human detection algorithm based on YOLOv5, during a typical conference, the peak number of people detected by the entrance sensor was 15, and the peak number detected by the central sensor was 13. A sliding window length is used. A data smoothing method for frames is used to calculate the real-time pedestrian flow statistics parameter vector. During the meeting, the number of people present was no less than 8 for 50% of the time periods, effectively reflecting fluctuations in the room's population density. The population statistics are synchronized into the main control unit's structured database via an interface, providing quantitative support for subsequent load demand prediction, comfort adjustments, and energy-saving strategies. Actual testing shows that the error rate of the population parameters after identity deduplication is less than 5%, significantly better than traditional single-point laser counting solutions. The dynamic output of these statistical parameters provides high-resolution time-series data for load prediction, achieving a high-precision closed-loop identification of room usage scenarios.

[0064] S1.6: Calls the meteorological data capture module through the external meteorological information interface, integrates external meteorological parameters such as temperature, humidity, wind speed, and air pressure, realizes the structured input of external meteorological data, and provides boundary conditions for load forecasting and energy consumption analysis.

[0065] S1.7: Retrieve building orientation parameters such as room orientation and exterior wall attributes from the building information database, and integrate them with the aforementioned room temperature and humidity and meteorological data to generate a building orientation information table, which is used to improve the modeling of spatial distribution and environmental influencing factors.

[0066] Step S2: Perform time-series denoising, normalization, and outlier removal on the environmental parameter data to obtain a dataset suitable for cross-scene modeling. Specifically, this includes:

[0067] S2.1: The collected environmental parameter time series data (including room temperature and humidity, fan current, valve opening, equipment surface temperature, traffic statistics, external meteorological data and building orientation information) are processed by a multi-channel adaptive sliding filter algorithm to suppress time series noise, so as to reduce the information noise introduced by sensor fluctuations, short-term external disturbances and sudden data anomalies, and achieve smooth output of the filtered environmental parameter time series results.

[0068] The collected environmental parameter time-series data, including room temperature and humidity, fan current, valve opening degree, equipment surface temperature, room occupancy statistics, external meteorological data, and building orientation information, are used as inputs for this step.

[0069] A multi-channel adaptive sliding filter algorithm (parameters: number of channels K≥6, sliding window length W=5~15, adaptively adjusted according to the sampling frequency and noise characteristics of each parameter) is adopted to achieve noise suppression and signal smoothing of data from various acquisition channels. This algorithm independently sets the window length W for each acquisition channel. k By statistically analyzing historical sampling points within the current window, the impact of spikes and transient interference on time-series data can be reduced.

[0070] Furthermore, by analyzing the raw time-series data x detected from each channel... k (t) An adaptive weighted moving average filter is used to automatically adjust the distortion peak and smooth the data output. Specifically, for each data channel k, the following processing is performed:

[0071]

[0072] in The time series data is processed by adaptive weighted moving average filtering. For the k-th channel time The original sampling data, As an adaptive weighting coefficient, the weights are dynamically adjusted based on the data fluctuation, autocorrelation, and pre-defined channel characteristics within the window.

[0073] Furthermore, an outlier sensitivity suppression algorithm is used to adaptively reduce the weight of abruptly changing sample points (such as outliers with jumps greater than 3 times the mean square error) during the filtering process, avoiding disturbances to the smooth output caused by sudden anomalies. Within the continuous sampling period, a historical noise distribution modeling method is employed, based on the historical fluctuation standard deviation σ of the acquisition channel. k The weight distribution of the sliding window is dynamically updated, so that the filtering results can balance sensitivity and robustness.

[0074] Furthermore, the time series results after multi-channel filtering are periodically checked at the global level. A redundant channel cross-validation method is used to perform consistency analysis on parameter groups with strong spatial correlation (such as multiple points of temperature and humidity, and fan current at different locations). If abnormal distribution is found, local encryption smoothing is immediately triggered to achieve distribution optimization under sudden interference.

[0075] Through the above multi-step chain derivation, the original multi-channel environmental parameter time series data are transformed into high-quality time series parameters that have undergone noise suppression, sudden anomaly buffering, and smooth output. This provides a stable data foundation for subsequent normalization, consistency, and outlier removal processing, and ensures data quality for cross-scenario modeling.

[0076] For example, in a standard 20 m² conference room environment, six sensor channels (temperature, humidity, fan current, valve opening, equipment surface temperature, and people flow statistics) are configured, with a data acquisition interval of 30 seconds. For the temperature and humidity channels, the sliding window length is set to W=7, and the initial weights are equally distributed. The temperature data sequence at a certain moment is [23.6, 23.7, 23.8, 23.7, 26.5, 23.7, 23.8], where the 5th value is significantly higher than the mean due to a momentary sensor anomaly. After setting a 3σ threshold (23.7±3×0.1), the weight of the outlier automatically decreases to 0.1, while the weights of other normal points are each 1. After weighted moving average, the filtered output is 23.74°C, effectively suppressing sudden anomalies. The fan current channel uses W=5, with a historical standard deviation of 0.02A. During periods of sudden anomalies, the global weight factor decreases to 0.5, achieving stable data output during the fan's operation period. After full-channel processing, the noise variance of the original data was reduced by 60%, and the false detection rate of spots in the pedestrian traffic channel decreased to less than 3%. Through this step, the original environmental parameters of multiple channels all output smooth, anomaly-free, and high-confidence time-series results, laying a solid foundation for subsequent normalization and outlier detection.

[0077] S2.2: For the time series results of environmental parameters after sliding filtering, perform consistent scale mapping on each channel based on professional normalization operators (such as Z-score normalization or Min-Max normalization) to map each environmental parameter data to a unified dimension range, eliminate scale inconsistencies caused by differences in equipment models, layouts and the output range of acquisition hardware, and obtain a normalized environmental parameter matrix.

[0078] S2.3: For the normalized environmental parameter matrix, outlier removal processing of time series data is performed based on joint confidence interval detection and moving window statistical adaptive threshold algorithm. Abnormal sample points caused by acquisition device failure, outlier peaks and extreme external shocks are identified and removed to generate a high-confidence environmental parameter usable dataset without outliers, so as to ensure the rationality of data distribution and the robustness of subsequent feature extraction.

[0079] Based on the normalized environmental parameter matrix, a joint confidence interval detection method (parameter setting: confidence level α=0.95) is used to extract the statistical features (such as mean μ) of the time series data of each channel. k Standard deviation σ k ), and calculate the current sliding time window W for each channel. n The upper and lower confidence limits within the range form a dynamic detection threshold interval.

[0080] Furthermore, an adaptive thresholding algorithm based on the sliding window statistical method (sliding window length L) is used. w =50~100, adjusted according to the data sampling frequency), for each time series data point x k (t) Perform real-time in-window statistics to obtain the corresponding expected range:

[0081] [ , ]

[0082] in, and The k-th channel is in the sliding window W n Mean and standard deviation within, is the quantile coefficient under a normal distribution.

[0083] A joint decision-making mechanism is adopted to normalize the matrix n. k Data points falling outside the dynamic confidence interval in (t) are marked as anomaly candidate points. The anomaly candidate points of multiple channels such as valve opening and fan current are linked and cross-compared to aggregate collaborative anomaly features such as equipment category, spatial region, and data category to enhance the robustness of anomaly detection.

[0084] Furthermore, for abnormal candidate points, continuous peak identification and extreme external shock detection methods are adopted to eliminate extreme samples caused by short-term disturbances (such as sudden power outages or transient equipment switching), and combined with intraday periodic pattern analysis, to prevent normal periodic fluctuations from being misjudged as abnormal.

[0085] Based on the anomaly tolerance configured by the data acquisition device administrator (e.g., the allowable percentage of extreme fluctuations is <0.5%), after multiple rounds of screening, data distortion points caused by data acquisition hardware failure, long-term communication disconnection, strong sensor interference, etc., are completely eliminated, and the remaining high-confidence environmental parameter data points are merged into a usable environmental parameter dataset without anomalies.

[0086] Using the aforementioned joint confidence interval detection and adaptive moving window outlier removal algorithm, the data matrix n normalized in the previous step is... k (t) is transformed into a high-confidence dataset without outliers, thus ensuring the rationality of the environmental data distribution and the robustness of subsequent feature extraction.

[0087] For example, in a 25 m² office, five channels are configured to monitor temperature and humidity, fan current, valve opening, equipment surface temperature, and people flow. The sampling period is 30 seconds, and after long-term operation, a normalized matrix n is generated. k (t), totaling 2880 points / day. The confidence interval method was set with α=0.95. For the room temperature channel, the daily average value μ T =23.65, standard deviation σ T =0.46, through the sliding window (L w According to statistics (=100), its dynamic confidence bound range is [22.75, 24.55]. A total of 14 outliers were detected in the actual historical data, distributed during peak morning and evening equipment switching times and sensor restart periods. For the wind turbine current channel, through cross-comparison of equipment types and peak value identification, 5 significantly abnormal peak values ​​were further removed. Ultimately, the total outlier rate for all channels was <0.5%, outputting 2870 high-confidence environmental parameter data points, meeting the robust data input requirements for subsequent multivariate feature extraction and load prediction models. Application results show that after outlier removal, the subsequent modeling MSE of the dataset decreased by approximately 22%, the model generalization error convergence was significantly accelerated, and the system's ability to perceive weak features and mutation risks was significantly improved.

[0088] S2.4: Based on the available dataset of high-confidence environmental parameters without outliers, a multi-source hierarchical missing value completion algorithm (such as alternating between K-nearest neighbor interpolation and historical working condition playback) is adopted to fill in the missing data collection window intervals caused by short-term communication interruptions or sensor failures in certain time periods. The complete time series set of environmental parameters after full-dimensional completion is output to ensure that downstream modeling tasks obtain continuous and complete information input.

[0089] For high-confidence environmental parameters without outliers, missing value completion processing can be performed on the dataset to address incomplete time-series data caused by factors such as short-term failure of fan coil unit sensors and communication interruptions.

[0090] A multi-source hierarchical missing value completion algorithm is adopted (parameters: priority sequence is K-nearest neighbor interpolation K=5, historical working condition playback window W). r =24h), to achieve preliminary estimation and filling of the missing intervals detected by all acquisition channels (such as temperature and humidity, fan current, valve opening, etc.).

[0091] Furthermore, using the K-nearest neighbor interpolation method, with the K nearest non-missing time points of similar data as references, the similarity of samples is scored according to Euclidean distance or cosine similarity, and the data values ​​of missing time points are estimated using a weighted average method:

[0092]

[0093] in, For the missing data points that need to be filled, The first K reference non-missing samples were selected. This represents the normalized weight determined by distance or similarity.

[0094] Furthermore, for missing intervals over long periods or spanning multiple sampling periods, a historical operating condition playback interpolation method is used, with past W... r The existing operating condition sequences corresponding to the window time period are synchronized and aligned. By collecting time series segments with the same time and parameters from historical operating conditions, the average playback is performed to fill the gaps, enhancing the temporal continuity and physical rationality of the filled gaps.

[0095] Furthermore, in cases where multi-channel synchronization is missing, a hierarchical priority mechanism is used to prioritize filling the core acquisition channels (such as temperature and humidity, fan current channels) and then fill the auxiliary channels (such as pedestrian flow, weather, building direction) to ensure the integrity and trend consistency of the main control parameter sequence.

[0096] Furthermore, for all newly generated or supplemented data intervals, a consistency constraint check strategy is adopted, such as sliding window smoothness check, time series jump detection, and physical upper and lower bound verification, to screen whether the interpolation results cause discontinuity of the sequence, excessive jump, or physical unreasonableness. Abnormal situations are fed back for secondary correction until they conform to the global data distribution and the physical attributes of system operation.

[0097] By using a multi-source hierarchical missing value completion algorithm, datasets without anomalies but with missing values ​​are processed into complete time-series sets of environmental parameters with full-dimensional completion. This meets the requirements of subsequent modeling tasks for continuous, full, and high-confidence data, and provides a solid information foundation for cross-scene modeling, feature engineering, and strategy optimization.

[0098] For example, in a 700 m² open office, temperature, humidity, fan current, and valve opening were collected every 30 seconds, while pedestrian traffic and meteorological data were collected every 60 seconds. During actual deployment, network switching caused single-minute gaps in the temperature and humidity data (2 minutes), single-minute gaps in the fan current data (5 minutes), and single-minute gaps in the valve opening data (1 minute, scattered throughout the night), resulting in a total gap rate of approximately 0.5%. For single-point short-term gaps within 30 seconds, the K-nearest neighbor interpolation algorithm (K=5) was used, and valid data within 2.5 minutes before and after the gaps were weighted to fill in the gaps, ensuring the mean square error of the interpolated sequence was less than 0.03. For intervals of 5 minutes or more, historical operating condition sequences from the past 24 hours under energy-saving mode were used for replay interpolation, ensuring the goodness of fit (R²) between the filled interval's trend and the actual workday curve was greater than 0.92. All filled intervals passed the sliding window consistency test (window L=7), with no significant abrupt changes or discontinuous jumps. The final output is a complete time series matrix with consistent length, full content, and physical constraint verification, which saves about 30% of the anomaly screening time for downstream feature extraction and prediction modeling, and reduces the system's energy consumption prediction model MAE by 4.5%.

[0099] S2.5: After the environmental parameters are fully completed, the complete time series set is used to perform time series data alignment of the multi-channel acquisition path using time synchronization alignment algorithms (such as multi-channel dynamic time warping DTW) and master clock multi-source alignment technology. This unifies the time label and sampling step size, realizes the synchronization and coordination of parameters from different sources, and outputs a high-quality structured dataset with guaranteed time series consistency, which serves as the standard input for subsequent cross-scene modeling and knowledge transfer.

[0100] Step S3: Based on the preprocessed dataset, a feature extraction process integrating multi-source information is constructed for different room types, area attributes, and equipment aging status. This process extracts load change features, comfort indicators, and energy consumption trend features, and generates multi-dimensional operating condition labels. For example... Figure 2 As shown, it specifically includes:

[0101] S3.1: For the preprocessed dataset, establish a multi-source information grouping index based on room type, area attributes and equipment aging status. Use a feature hierarchical mapping algorithm to archive the environmental parameter data according to the group to obtain data partitions subdivided by application scenario, laying a prior data foundation for subsequent working condition similarity matching and specific strategy optimization.

[0102] For the complete time series set of environmental parameters after full-dimensional completion, a multi-source information grouping index is established using room type, area attribute and equipment aging status as grouping variables to achieve structured archiving of data sources.

[0103] A feature-based hierarchical mapping algorithm (parameters: grouping dimensions cover room function type, space partition number, sensor device identifier, and device health interval) is used to classify environmental parameter data step by step, and aggregate and archive data from similar rooms, similar areas, or health intervals.

[0104] Furthermore, a semantic tag mapping mechanism is adopted to map the above grouped data to a preset application scenario tag system (such as "meeting room-west area-aging"), generating data partitions with scenario-side tags.

[0105] By using a data integrity verification algorithm, the key parameters (temperature, humidity, fan current, etc.) in each data partition are verified for coverage, and partitions with a collection volume lower than the set threshold are filtered out to ensure that each application scenario partition has a sufficient number of usable data entries.

[0106] Furthermore, using zonal characteristic frequency analysis, descriptive statistics are performed on the distribution of load characteristic parameters, comfort indexes, and energy consumption trends in each data zonal, extracting statistical characteristic indicators for each zonal, and providing data support for subsequent operating condition similarity retrieval and personalized strategy optimization.

[0107] By using a feature-based hierarchical mapping algorithm, the full set of environmental parameter data is structured and transformed into fine-grained data partitions that meet the requirements of grouping labels, statistical properties, and physical consistency. This enables precise feature hierarchical mapping for different application scenarios, laying a high-confidence prior data foundation for subsequent multivariate feature extraction and scenario-based modeling tasks.

[0108] For example, in the air conditioning management system of a Grade A office building, environmental parameter data for 600 independent rooms were collected simultaneously, involving three typical room types (standard offices, meeting rooms, and server rooms). Each type of room was divided into four spatial areas: east, west, south, and north. The fan coil units in each room were categorized into three health levels based on their actual service life: "brand new (0-3 years)," "nearly new (3-7 years)," and "aged (>7 years)." For the collected full data, a feature-based hierarchical mapping algorithm was used. First, a first-level index was constructed based on the room type field. Then, it was further subdivided based on the spatial area number. Finally, the health range of the fan coil units was introduced as a third-level index variable, merging all room environmental parameters into 36 independent partitions (3 room types × 4 areas × 3 health levels). In the meeting room-west area-aged partition, the data volume was 19,820 records, covering all six parameters including temperature, humidity, and fan current, achieving a coverage rate of 99.2%. One partition with abnormal data volume was removed using a data integrity check algorithm. Statistical analysis of the characteristic frequencies of the data within each partition revealed a root mean square deviation of 0.43℃ for temperature fluctuations, an average fan current of 0.22A, and a high deviation in the surface temperature distribution of equipment, indicating a high maintenance load and significant signs of aging in the rooms within that partition. Ultimately, the data from each partition possesses the characteristics of scenario-based scripted applications, outputting multi-source information partitions indexed by a three-dimensional structured label of room type-region-health level. This provides high-confidence grouped input data for subsequent specialized modeling and global transfer learning strategies, supporting a 30% reduction in the average value of manual inspections and a 15% improvement in the accuracy of risk warnings for aging zones.

[0109] S3.2: Using multi-source information zoning environmental parameters as input, a multivariate dynamic load modeling algorithm is adopted to extract key load change features of rooms, including time-series changes in cooling / heating loads, fan coil unit power fluctuations, and valve response amplitudes, and outputs a load change feature parameter sequence to provide high-dimensional feature input for energy scheduling and load risk prediction.

[0110] S3.3: Based on the same multi-source information partitioned environmental parameters, and using indoor comfort assessment index system (such as PMV, PPD index and other human comfort algorithms), calculate the comfort index characteristics such as temperature and humidity, wind speed, and surface temperature difference, and generate an indoor comfort characteristic parameter table for each time period, providing input boundaries for the target weight of the control strategy.

[0111] S3.4: Based on the aforementioned load change characteristics and comfort index characteristics, a time-series energy consumption decomposition algorithm and sliding window statistical analysis are used to extract features from energy consumption-related parameters such as fan current and valve opening, generating an energy consumption trend feature vector, which provides core constraint data for subsequent energy consumption prediction and energy efficiency optimization models.

[0112] S3.5: Using the obtained load change characteristic parameters, comfort characteristic parameters, and energy consumption trend characteristic vector as input, the data is mapped to generate multi-dimensional operating condition labels through multi-dimensional operating condition label generation rules. The label content includes information such as room partition number, current equipment health status, load type, comfort level, and energy consumption level, providing a structured scenario description for cross-domain knowledge transfer and strategy optimization models.

[0113] Using the obtained load change characteristic parameters, comfort characteristic parameters, and energy consumption trend characteristic vector as input, a multi-dimensional working condition label generation rule algorithm (parameters: label dimensions are defined as partition number, health status, load type, comfort level, and energy consumption level) is adopted to realize the structured working condition label mapping function.

[0114] Furthermore, by associating the tag mapping table, and based on the spatial partition number, equipment aging status, and partition function type, the load change characteristic sequence is linked with comfort indicators and energy consumption trend values ​​at the scene level. Using tag encoding algorithms (such as independent hot coding and multi-level hierarchical mapping technology), elements such as spatial partition number P, equipment health H, load type L, comfort level S, and energy consumption level E are encoded dimension-by-dimensionally to generate a preliminary tag matrix T.

[0115]

[0116] Where N is the number of samples and D is the label dimension. P is taken from the room zoning coding table, H is based on the equipment operation health evaluation model (parameters refer to mean square error, absolute deviation and abnormal statistical characteristics), L is based on the time series feature clustering results to calibrate the cold / heat / mixed load type, S obtains the comfort level range through PMV and PPD threshold mapping, and E is quantified to the standard classification using energy consumption trend characteristics.

[0117] Furthermore, the initial label matrix T is deduplicated and normalized using label consistency verification algorithms (such as redundant label deduplication, confidence level judgment, and label conflict decomposition) to eliminate abnormal conflicts or fuzzy duplicate labels, ensuring that each data instance corresponds to a unique and multidimensionally consistent label combination.

[0118] Furthermore, by employing tag synthesis and multidimensional normalization techniques, the multidimensional encoding of tags is compressed into a standardized tag vector T*, facilitating subsequent embedded processing in knowledge transfer and strategy optimization models. This tag vector, in the form of sparse hot encoding or vectorized embedding, possesses a complete scene description capability, including partition positioning, health status differentiation, load characteristics, comfort goals, and energy consumption goals.

[0119] By using multi-dimensional operating condition label generation rules, the original information of the above-mentioned load change characteristic parameters, comfort characteristic parameters, and energy consumption trend characteristic vectors are transformed into a multi-dimensional operating condition label structure, realizing an accurate and transferable expression of complex room-equipment-energy consumption scenarios, and providing high-confidence structured input data for cross-domain knowledge transfer and strategy optimization of reinforcement learning models.

[0120] For example, in a typical office building, 19,820 scenario samples from the conference room-west area-aging equipment zone were used as the input. The load change characteristic sequence (average cooling load 3.2kW, fluctuation variance 0.7kW^2), comfort parameter table (PMV=0.4, PPD=8%), and energy consumption trend characteristic vector (average daily maximum power 0.78kW, energy consumption growth rate 1.5% / h) were used as input.

[0121] The tag encoding algorithm is adopted, the partition number P is specified as

[03] , the health level H is evaluated as [aging (2)], the load type L is clustered as [cooling load (1)], the comfort level S is quantified as [relatively comfortable (2)], and the energy consumption level E is divided into [high (3)]. The tag matrix T is: [03,2,1,2,3]. There is no redundancy or conflict in the tag redundancy detection. The tag vector is directly normalized and synthesized into a five-dimensional standardized tag vector T*=[0.03, 0.66, 0.33, 0.66, 1.00].

[0122] Finally, this embodiment outputs a multi-dimensional operating condition label T*, which serves as the batch input for subsequent transfer learning models and policy network modules. This supports the system's knowledge transfer and energy efficiency priority control mechanism for extreme scenarios such as aging equipment, high load, high energy consumption, and local comfort. The measured accuracy of automatic label generation reaches 99.8%, effectively improving the global scene recognition and parameter transfer generalization performance.

[0123] Step S4: Input the multi-dimensional operating condition labels into the cross-domain knowledge transfer model, and use the energy consumption and comfort adjustment parameters learned in the known scenario to perform knowledge transfer and parameter pre-initialization to obtain the parameter priors of the current target environment. Figure 3 As shown, it specifically includes:

[0124] S4.1: For the multi-dimensional working condition labels generated during the multi-source information fusion process, based on the regional attributes, room type and equipment aging status corresponding to the labels, the label feature vector is normalized using a label standardization algorithm to obtain label feature vectors suitable for cross-domain migration, thus preparing for the standardization of the subsequent knowledge transfer model input.

[0125] Cross-domain knowledge transfer models include: domain adaptation models, knowledge distillation models, feature alignment models, parameter sharing models, federated learning models, federated cross-domain recommendation models, transfer reinforcement learning models, relation transfer models, and so on.

[0126] The multi-dimensional operating condition labels generated by the multi-source information fusion module include a structured label matrix T* after processing in the previous stage (feature hierarchical mapping, label encoding and consistency verification). Each label instance corresponds to a unique room partition number, equipment health status, load type, comfort level and energy consumption level code.

[0127] A label standardization algorithm is adopted, and for each dimension of the label matrix T* (including spatial number, health level, load type, comfort level, and energy consumption level), a normalized mapping interval (parameter: standard normalization interval [0,1]) is configured based on the information distribution characteristics and the consistency requirements of the target cross-domain scenario. For numerical labels (such as equipment health and energy consumption level), the algorithm uses Min-Max normalization.

[0128]

[0129] in, These are the specific tag elements after normalization. For the specific tag elements to be normalized, , These are the minimum and maximum values ​​of the label parameter in the entire dataset.

[0130] For label dimensions with discrete attributes (such as room zoning, load type, and comfort level), a hybrid strategy of one-hot encoding and normalized heat encoding is adopted.

[0131]

[0132] in, Let i be the one-hot encoded value of the i-th sample in the j-th label dimension. for The corresponding normalized one-hot encoded value, Given the total number of unique values ​​for this label dimension, output a normalized encoded vector.

[0133] Furthermore, through label distribution visualization and statistical consistency testing, the Kolmogorov-Smirnov test or Kullback-Leibler divergence analysis is performed on the label feature vector distribution of the batch normalized output to ensure that the label normalization results of different room types, regional attributes and aging states have cross-domain scene distribution consistency, providing a standard template for subsequent cross-domain transfer learning model input.

[0134] Through the above label standardization algorithm, multi-dimensional working condition labels are converted into standardized label feature vectors in batches, realizing unified scale mapping and structured expression among multi-source labels, and preparing the optimal feature foundation for the input of cross-domain knowledge transfer models.

[0135] For example, in the aforementioned office building conference room-west area-aging equipment scenario, the input label matrix T is [03,2,1,2,3]. After Min-Max normalization, the room number P is normalized to 0.03, the equipment health H to 0.66, the load type L to 0.33, the comfort level S to 0.66, and the energy consumption level E to 1.00. L2 norm normalization is then performed on the label vector, outputting the standard label feature vector T. * =[0.021,0.464,0.232,0.464,0.704]. Normalization was performed on the 19820 label vectors. The distribution consistency test result showed a Kolmogorov-Smirnov distance of less than 0.07, satisfying the cross-domain input uniformity constraint. Finally, the standardized label feature vectors output from this step were input into the cross-domain transfer learning model to support knowledge parameter mapping. This enabled efficient transfer of energy efficiency strategies and comfort goals under different room and equipment health states, supporting a more than 20% improvement in strategy tuning convergence speed in real-world scenarios. The label conversion accuracy reached 99.8% in energy-saving application test scenarios.

[0136] S4.2: Using the standardized label feature vector as input and the energy consumption characteristic parameters and comfort adjustment parameters accumulated from historical known scenarios as the source domain, a domain adaptive transfer learning algorithm is adopted. Under the transfer learning framework, the implicit correlation features between different labels are compared. Through parameter mapping and distribution matching, the set of prior adjustment parameters that are highly correlated with the current target label is extracted.

[0137] Using the standardized label feature vector T output by the label standardization algorithm as input, this model is fed into a cross-domain knowledge transfer learning model to achieve efficient transfer matching of parameters between the source and target domains. A domain-adaptive transfer learning algorithm (parameters: feature space consistency mapping, distribution matching loss threshold, feature extraction kernel) is employed to jointly align energy consumption characteristic parameters and comfort adjustment parameters within the source domain (historically known scenarios) with the label feature space of the current target environment. Through hidden layer feature embedding, the label feature vector T is spatially projected and reconstructed using principal components with the source domain parameter distribution, generating a parameter mapping feature tensor.

[0138] Furthermore, the maximum mean difference (MMD) distribution matching algorithm (parameters: Gaussian kernel for kernel type, feature dimension adjustment coefficient β, and sample batch normalization) is used to match the source domain parameter distribution P. s (X) and the distribution of the target label feature domain P t (X) Perform adaptive distribution convergence calculation and optimize the loss function.

[0139]

[0140] in, The value of the loss function. For kernel space mapping functions, For source domain kernel space mapping function, X is the kernel space mapping function of the target label feature domain. s X represents the sample features sampled from the source domain. t For sample features sampled from the label feature domain, For the expectation operator, by reducing This achieves the alignment of the distribution of the source domain and the target label feature domain.

[0141] Furthermore, a parameter-level mapping and feature relevance screening algorithm (parameters: label similarity score threshold η, feature interaction weight λ) is employed. In the source domain empirical parameter library (coverage energy consumption curve response coefficient, comfort target adjustment weight, dynamic threshold setting), historical parameters are screened item by item according to feature relevance alignment values. The mapping matrix W is used to achieve the optimal projection of the historical parameters into the label space, forming a set of highly relevance candidate prior adjustment parameters. The formula is:

[0142]

[0143] in, For the complete set of source domain parameters, This is a similarity measurement function between the label vector and the parameter vector (cosine similarity and Euclidean distance are both applicable). The mapping matrix projects the parameters from the original space to the label space or feature space, making the parameters comparable to the labels. For a single parameter vector, The projection vector of the parameters in the label space; The target label vector; The goal is to select the subset of parameters that maximizes similarity from the complete set of parameters in the source domain.

[0144] Furthermore, an implicit feature residual screening algorithm (parameters: error tolerance interval δ, feature contribution ranking) is used to process the candidate parameter set obtained from the mapping. Residual zeroing is performed, removing historical parameter components that cannot adapt to the high-weight features of the target working condition, and retaining only the subset of prior parameters that are highly coupled with the current label feature vector T. Through the above processing chain, the domain adaptive transfer learning algorithm driven by standardized label feature vectors, distribution matching, and parameter screening techniques are organically combined to obtain a priori adjustment parameter set adapted to the current target label, thereby achieving a systematic improvement in parameter adaptation accuracy and convergence speed during knowledge transfer. For example, in a Grade A office building conference room-west zone-aging equipment environment, the standardized label feature vector T=[0.021,0.464,0.232,0.464,0.704] is input. Referring to the K-nearest neighbor distance weighting strategy, energy consumption curve response parameters and comfort control parameters with an Euclidean distance less than 0.12 in the source domain's historical scenarios are selected to form a preliminary source domain parameter subset (67 groups in total). Through maximum mean difference (MMD) distribution matching, a Gaussian kernel bandwidth parameter σ=0.5 is selected. After batch normalization, the MMD loss between the source and target domains can be reduced to 0.008. The correlation screening parameter η is set to 0.85, and parameters with a cosine similarity higher than 0.85 are screened out as candidates. A total of 41 groups were identified. Based on feature residual screening, low-coupling parameter groups with δ>0.05 were removed, resulting in 28 highly correlated prior adjustment parameters. In actual system deployment, the cold start strategy improved the pre-response capability by 19.2% during the transfer learning phase and the convergence speed of energy efficiency optimization by 23% in 12 hours. The prior parameters can effectively support the risk suppression and energy efficiency improvement goals under the range of aging equipment, high load, and high energy consumption.

[0145] S4.3: Based on the prior adjustment parameter set and target label feature vector obtained above, a transfer regularization constraint mechanism is adopted to re-tune the prior adjustment parameters and adaptively adjust the weights to eliminate potential distribution deviations between the source domain and the target environment in terms of energy consumption control and comfort indicators, thereby outputting prior parameters that are adapted to the current target environment.

[0146] S4.4: Based on the obtained parameter priors, combined with the label feature vector and the dynamic operating conditions of the room, parameter clustering and sensitivity analysis algorithms are used to cluster and archive each indicator (such as cooling load adjustment sensitivity, wind speed efficiency threshold, and prediction time window) within the parameter priors and to screen key parameters, so as to obtain a parameter prior set with more universality and generalization ability.

[0147] S4.5: The prior set of parameters obtained from the final screening is input into the subsequent neural network time series prediction module and digital twin simulation unit to initialize model parameters and simulation scenario settings, realize system-level dynamic adaptive initialization and cross-scenario control capability improvement, and provide highly reliable basic data for the optimization strategy of the reinforcement learning controller.

[0148] Step S5: Based on a neural network time-series prediction model, combined with internal sensor data and external dynamic data, the load trend and equipment anomaly probability of the room within a future preset time window are predicted, and the load mutation risk value and anomaly tendency characteristics are output. Specifically, this includes:

[0149] S5.1: The internal sensor data (including room temperature and humidity sensor data, fan current signal, valve opening signal, and equipment surface temperature signal) and external dynamic data (including room traffic statistics, meteorological data, and building orientation information) are synchronized and organized. A time series resampling and alignment algorithm is used to generate a multi-source feature time series matrix under a unified time reference to ensure that the data is consistently available in the time series prediction model.

[0150] A unified time base and sampling frequency are set for internal sensor data (including room temperature and humidity sensor data, fan current signal, valve opening signal, and equipment surface temperature signal) and external dynamic data (including room traffic statistics, meteorological data, and building orientation information) to form multi-source raw data input.

[0151] The time-stamp alignment method (parameters: master clock reference, NTP or PTP synchronization protocol) is used to achieve time synchronization of raw data streams from different sensor nodes, mapping various types of data to a standard time axis and eliminating cross-device sampling delay and clock drift.

[0152] Furthermore, by using a time-series resampling algorithm (parameters: target sampling period Δt, resampling interval selection: linear interpolation or spline interpolation), resampling is performed on individual channels with inconsistent frequencies or missing data acquisition frames to fill in all the data points that should be on the time axis, unify the data sampling frequency, and obtain the equally spaced time-series data sequence for each channel.

[0153] Furthermore, a multi-source data path aggregation algorithm (parameters: multi-channel identifier, spatial location encoding, data validity mask) is adopted to aggregate the processed time series data, construct a multi-dimensional feature column with unified channel number, remove invalid or abnormal sampling frames, and summarize into a synchronous data set of all features at time t.

[0154] The index mapping and hash archiving method (parameters: sample stride, data window length) is used to transform the multi-source aggregated data sequence into a fragmented structure according to the specified stride, forming a standardized time series data block containing all feature columns, which facilitates subsequent windowed feature extraction and convolution operations.

[0155] Through the above synchronization, resampling, and aggregation processes, a multi-source feature time series matrix under a unified time reference is output. , where N is the number of sampling steps and d is the comprehensive feature dimension, to achieve consistent input, collaborative modeling and efficient integration of cross-source features in multi-objective time series prediction models.

[0156] For example, in the scenario of the conference room in the west wing of an office building, the sampling period for room temperature and humidity sensors is 2 seconds, for fan current / valve opening / surface temperature is 5 seconds, for pedestrian flow camera data is 10 seconds, and for meteorological and building orientation information, the update period is 60 seconds. All nodes calibrate their master clocks via the NTP protocol, using a 2-second sampling step. Fan and valve values ​​are padded with linear interpolation to complete the step size, while meteorological and orientation parameters are padded with nearest neighbors to align short periods. Through hash channel aggregation, the data buffer outputs 30 frames of structured data every 60 seconds, each frame including 14-dimensional channel features (temperature and humidity 4, fan 2, valve 2, surface temperature 2, pedestrian flow 1, meteorological 2, orientation 1). In the fault simulation scenario, the maximum sampling delay per period is less than 0.7 seconds, and the residual clock drift error is less than 0.001%. The final output... The time series matrix is It achieves 100% step-size alignment with no missing data, realizes consistent and accurate synchronization of all internal and external data and cross-source instruction collaborative input, and stably supports windowed dynamic modeling of multi-objective neural network backends.

[0157] S5.2: Based on the multi-source feature time series matrix, feature engineering algorithms (such as sliding window statistics, change point detection and autocorrelation analysis) are used to dynamically extract features from the time series signal to obtain the load change characteristics, comfort index and energy consumption trend characteristics of the fan coil unit terminal, so as to form a multi-parameter input tensor suitable for further time series modeling.

[0158] S5.3: Input multi-parameter input tensors into a neural network time series prediction model, and use a long short-term memory network (LSTM) and attention mechanism to encode load trends and equipment anomaly risks in parallel, capturing the time dependency of load cascading mutation patterns and equipment health hazards.

[0159] S5.4: Perform multi-objective loss function optimization on the time series prediction results. Based on the energy consumption prediction error, comfort evaluation index error and equipment abnormal risk detection accuracy, perform joint weight optimization iteration to output the load mutation risk value and abnormal tendency characteristics within the future preset time window.

[0160] S5.5: Dynamically aggregate and attribution analyze the load mutation risk value and abnormal tendency characteristics. Employ mutation threshold detection and anomaly probability attribution algorithms to identify high-risk moments and corresponding dominant features, providing decision-making suggestions for prioritizing digital twin simulation and reinforcement learning strategies.

[0161] Step S6: Construct a real-time digital twin simulation unit. Based on the prior parameters and the mutation risk value, perform virtual calculations on the execution consequences of various adjustment strategies, predict the impact of fan coil unit adjustment actions on temperature control response and energy consumption, and determine the optimal adjustment timing window and activation threshold. Specifically, this includes:

[0162] S6.1: Initialize the operating conditions for the parameter priors and mutation risk values. Based on the parameter priors and load mutation risk values ​​output by the cross-domain knowledge transfer model and the multi-objective neural network time series prediction module, complete the initial parameter configuration of the simulation environment of the real-time digital twin simulation unit to ensure that the simulation experiment accurately reflects the current operating conditions of the central air conditioning fan coil unit terminal.

[0163] S6.2: By combining physical modeling and data-driven modeling methods, multi-dimensional modeling of fan coil unit regulation strategy variables is performed. By establishing a digital twin simulation network covering the thermodynamic characteristics of the main unit and terminal units, ventilation efficiency, and dynamic load response parameters, the terminal response behavior of fan coil units under multiple operating conditions is modeled, providing a reliable model foundation for virtual calculations.

[0164] S6.3: Based on the input of initialization parameters and adjustment strategy variables under operating conditions, perform batch virtual calculations on the execution consequences of different adjustment strategies (including heating, cooling, and fan speed adjustment) of fan coil units. Employ real-time numerical simulation and state feedback mechanisms to output dynamic prediction data of room temperature changes, energy consumption changes, and indoor comfort indicators for each strategy.

[0165] S6.4: Perform multi-objective time-series evaluation on the temperature response curve, delay threshold, and energy consumption curve output by the virtual calculation of the strategy actions. Use the operating condition discrimination algorithm to standardize and quantify the benefit-cost function of the consequences of each regulation strategy, and screen the regulation time sequence window that meets the balance of energy efficiency, hysteresis tolerance, and comfort, so as to provide decision-making reference for the subsequent determination of the strategy activation threshold.

[0166] S6.5: Based on the timing evaluation results and simulation output data, a decision optimization algorithm is used to calculate the optimal timing window and activation threshold of the fan coil unit terminal adjustment strategy, optimize and determine the set of control actions that can be triggered in advance, and generate the optimal adjustment timing window and activation threshold parameter set that can be dynamically called by the reinforcement learning controller to achieve efficient strategy linkage with the physical system.

[0167] The input consists of core operating condition evaluation results, such as temperature response curves, delay thresholds, and energy consumption curves, obtained through multi-objective time-series evaluation and digital twin simulation, as well as simulation output data including initial parameters and strategy variables of the actual operating conditions.

[0168] A piecewise dynamic programming decision optimization algorithm (parameters: adjustment time interval, strategy action set, benefit-cost function) is adopted to perform stepwise optimality analysis on the effect of adjustment strategy at different times, realize multi-objective adaptive window search, and establish the optimal pairing relationship between strategy action and actual working condition response.

[0169] Furthermore, a multi-objective benefit-cost function quantification algorithm (parameters: expected energy consumption reduction ΔE, target comfort score ΔC, and adjustment delay acceptance threshold ΔT) is used to standardize and quantify each adjustment time window and action trigger point, employing the following comprehensive performance evaluation index:

[0170]

[0171] in, For a moment Next action The overall performance score, The change in energy consumption resulting from this action. To improve comfort, The response lag introduced to control the action, , , The weights of each objective, , , This is the normalized baseline value.

[0172] Furthermore, using a sliding window sensitivity analysis algorithm (parameters: window length L, time delay step size δt), local extreme value segments are searched among all feasible timing schemes to determine the optimal adjustment timing window [t]. s ,t e And solve for the corresponding activation threshold. The following equation is satisfied:

[0173]

[0174] Through the joint optimality search of the sliding window and action space described above, a set of control actions that achieves optimal energy efficiency and acceptable comfort under the target operating conditions is selected. , The optimal start time of the time window. This represents the end time of the optimal time window. The optimal activation threshold; The maximization operator represents selecting the optimal time window and threshold combination from all possible combinations. The largest group , The start time of the time window. The end time of the time window. The activation threshold; Let be the objective performance function, and comprehensively evaluate the performance at time t and threshold . At that time, the weighted performance of energy efficiency and comfort; As a constraint, it means that the performance at all times within the entire time window must meet certain requirements (such as comfort not being lower than a certain lower limit), rather than optimizing only a certain moment.

[0175] Furthermore, an action parallel grouping and archiving mechanism (parameters: room ID, policy label, priority hierarchy) is adopted to output the optimal adjustment timing window and activation threshold in a multi-layered structure according to room, partition and policy type, generating a set of parameters that the reinforcement learning controller can directly and dynamically call.

[0176] By using the aforementioned decision optimization algorithm and time sensitivity analysis method, the time evaluation and simulation output results are transformed into optimal adjustment time window and activation threshold parameters, enabling early triggering and intelligent linkage of the control actions at the end of the fan coil unit, and significantly improving the system's ability to respond proactively to sudden load changes and abnormal risks.

[0177] For example, in the scenario of the conference room in the west wing of the office building, the time window for digital twin simulation output is 120 seconds, the adjustment action set includes 3 fan speeds and 2 temperature settings, and the energy consumption baseline is... Wh, the comfort rating benchmark is The time delay reference is Set weights , , For each action 'a', the time 't' at each moment is calculated using the above formula. , Let be the objective performance function, and comprehensively evaluate the performance at time t and threshold . At that time, the weighted performance of energy efficiency and comfort, and the length of the sliding window. Step length Throughout the entire simulation time window, the system automatically filters out... , The optimal adjustment window for energy efficiency and comfort corresponds to the activation threshold. The high-speed fan is triggered 15 seconds in advance. The final output parameter set used for the reinforcement learning controller is {room ID, window [40, 65], instruction a^*=high speed, priority=1}. In actual deployment, the conference room can respond to sudden load changes in advance under conditions of sudden increase in people flow and external temperature rise, achieving a 4% reduction in energy consumption, a comfort decrease of less than 1% of the limit, and a reduction in adjustment response lag to within 8 seconds, significantly improving the system's strategic linkage and risk prevention capabilities.

[0178] Step S7: Using the optimal adjustment timing window and activation threshold as input, dynamically optimize the policy network in the reinforcement learning controller. Based on the three-element adaptive adjustment mechanism of "expected energy consumption benefit - policy execution lag - comfort tolerance", optimize the pre-execution control commands of the output fan coil unit. Specifically, this includes:

[0179] S7.1: Based on the optimal adjustment timing window and activation threshold output by the digital twin simulation unit, the timing window and activation threshold are used as inputs to the reinforcement learning controller to ensure that the policy network can integrate the dynamic inference results of the simulation unit on the load trend of the fan coil unit and the risk of equipment anomalies, and use them to adjust the control policy adaptation parameters at the current moment.

[0180] S7.2: The optimal adjustment time window, activation threshold, and load mutation risk value and abnormal tendency characteristics generated by the multi-objective neural network time-series prediction model are jointly processed. Through the adaptive weight allocation algorithm in the policy network structure, the weight parameters of the three-element performance indicators such as 'expected energy consumption benefit', 'policy execution lag', and 'comfort tolerance' are allocated in the decision-making process to achieve dynamic trade-offs in multi-objective optimization.

[0181] The optimal adjustment time window, activation threshold, and load mutation risk value and abnormal tendency characteristics generated by the multi-objective neural network time series prediction model are used to achieve structured input of different types of control parameters in a unified decision space by employing a multi-source heterogeneous data fusion method (parameters: time series window boundary, threshold parameter, load risk vector, and abnormal probability characteristics).

[0182] An adaptive weight allocation algorithm (parameters: expected energy consumption benefit target, policy execution lag tolerance, and comfort tolerance threshold) is adopted in the policy network to establish a three-element performance index mapping relationship for the fused parameter set. 'Expected energy consumption benefit', 'policy execution lag', and 'comfort tolerance' are embedded as independent performance dimensions into the policy network weight layer.

[0183] Furthermore, through a multi-objective optimization constraint mechanism, a joint loss function for the policy network performance objectives is defined, and the ternary performance index is dynamically weighted using the following weight allocation formula:

[0184]

[0185]

[0186] in, Weighted by expected energy consumption revenue, For comfort tolerance weight, The weights for strategy execution lag are calculated, and the sum of the three is 1. This is the overall performance score after weighting.

[0187] The change in energy consumption resulting from this action. To improve comfort, The response lag introduced to control the action, , , This is the normalized baseline value.

[0188] A dynamic weight adaptive allocation method is adopted (parameters: historical strategy execution effect backtracking, mutation risk detection probability, and regulation lag trend). Based on a periodic rolling window, the expected value, variance, and response sensitivity coefficient of each performance index are calculated, and the values ​​are dynamically adjusted according to the system status and predicted output under real-time operating conditions. , , .

[0189] Furthermore, a weight normalization mapping constraint (parameters: normalization benchmark, extreme value shearing) is adopted to project the weight components of each performance index into intervals, forcibly constraining the extreme value transitions of the weights caused by sudden changes in operating conditions, and avoiding nonlinear instability of the policy network.

[0190] Through the above chain derivation, a joint dynamic weight allocation for the optimal adjustment time window, activation threshold, and related prediction indicators is realized. The multi-objective optimization objective is continuously embedded into the policy network structure in the form of adjustable parameters, effectively supporting the dynamic optimization and action selection of the downstream policy network.

[0191] For example, taking the conference room in Area A of the office building as an example, the optimal adjustment timing window is [30s, 60s], the activation threshold is high wind speed, the load mutation risk value is 0.85, and the abnormal tendency characteristic is 0.15. Periodic backtesting based on the control data of the most recent 5 days shows that the historical cumulative energy consumption reduction averages 6.5%, comfort score fluctuations are controlled within ±2%, and the average response lag is 12 seconds. Initial weights are set. (Focus on energy conservation) (Moderately comfortable) (Moderate risk of lag) According to real-time monitoring, if the predicted probability of a sudden increase in load rises to 0.92, the system will automatically perform rolling upgrades. Reduced to 0.25 To minimize control latency, a certain level of comfort is sacrificed. After comprehensive weight allocation, the policy network dynamically outputs within this time window. The optimal action distribution provides the optimal parameters for subsequent value function estimation and action selector. Actual deployment results demonstrate that the energy consumption response curve and comfort score improved simultaneously, and the policy network exhibited the ability to adapt to and optimize decisions in the event of sudden loads and abnormal conditions.

[0192] S7.3: Using the above adaptive weight parameters, perform value iteration calculation on the value function estimation module in the policy network, so that the policy evaluation process fully considers the influence of the adjustment window (cause) obtained by simulation on the final control activation time and comfort (effect), and obtains the dynamically optimized control policy vector.

[0193] S7.4: Input the dynamically optimized control strategy vector into the action selector in the reinforcement learning controller, and adjust the action probability distribution based on the soft maximum strategy or ε-greedy strategy, thereby outputting a pre-execution control command for the fan coil unit that takes into account energy efficiency, time delay and comfort tolerance.

[0194] S7.5: Performs post-consistency verification on the output fan coil unit pre-execution control commands. Based on the current cycle prediction error and strategy execution delay feedback, it corrects the parameter update rate and action threshold in the strategy network in real time, ensuring that the strategy optimization in the next cycle can continuously self-evolve according to the historical causal chain, realizing the online adaptive and robust output of the central air conditioning fan coil unit reinforcement learning controller.

[0195] Step S8: According to the pre-execution control command, an operation command is issued to the central air conditioning fan coil unit terminal in real time to activate the heating, cooling, and fan speed adjustment commands in advance, so that the system can dynamically respond to predicted load changes or equipment abnormalities. Specifically, this includes:

[0196] S8.1: Based on the pre-execution control instructions obtained in the previous step, analyze the specific parameters in the control instructions, including the target wind speed, the desired temperature setpoint, and the execution priority label, to obtain a complete set of fan coil unit terminal operation commands, which are used to clarify the data assignment of each control entry point of the equipment.

[0197] S8.2: Determine the execution status of the fan coil unit terminal, obtain the current fan operation status, valve opening and closing status and on-site feedback of the terminal equipment, and verify the equipment linkage conditions through the equipment status synchronization mechanism to ensure that the time window for issuing operation commands matches the actual controllable status of the equipment, thereby optimizing the execution sequence for downstream actions.

[0198] S8.3: Employs field-level communication protocols (such as Modbus, BACnet, etc.) and uses encrypted command channels to send generated fan coil unit terminal operation commands in batches and according to priority to the corresponding fan coil unit controllers in each room in real time, ensuring full coverage of control commands and anti-interference capabilities.

[0199] S8.4: Implement command mapping on the fan coil unit controller side, and map the received operation commands into hardware action commands through the device driver layer to realize hardware-level trigger control of heating function, cooling function and fan speed level, and complete the physical action closed loop.

[0200] The set of operation commands at the fan coil unit terminal is used as input data, including target wind speed command, desired temperature set value and control priority parameters. It is required to complete the mapping from command to physical action on the fan coil unit controller side to ensure short-time response.

[0201] The instruction parsing protocol layer module (parameters: operation command format, function enumeration table, instruction decomposition rules) is used to classify and parse the received set of operation commands from the fan coil unit terminal, distinguish between multiple control functions such as heating, cooling and fan speed adjustment, and accurately extract key control parameters and action priority identifiers.

[0202] Furthermore, through the device driver mapping adapter (parameters: driver interface dictionary, hardware adaptation layer, control mapping table), the parsed logic commands are converted into corresponding underlying hardware action codes according to the driver layer mapping rules, forming level signals, PWM adjustment signals and fieldbus data formats compatible with the terminal execution device, ensuring smooth linkage between operation commands and the fan coil unit controller hardware control unit.

[0203] Furthermore, a multi-channel hardware scheduling mechanism (parameters: heating relay control port, cooling valve execution port, fan speed PWM signal channel, priority queue) is adopted to implement parallel or time-sharing scheduling of each functional control channel according to the priority order set by the operation command, so that heating, cooling and fan speed adjustment commands can be triggered independently or in combination, while preventing command conflicts and hardware-level deadlock.

[0204] Furthermore, through the status monitoring and action closed-loop detection module (parameters: action feedback loop, in-situ detection signal, timeout protection threshold), the action response signals of each physical execution component of the fan coil unit are collected in real time to determine whether each function operation has been executed accurately, and an alarm is issued in case of abnormality to ensure hardware-level safety and execution consistency.

[0205] Through the aforementioned drive mapping and closed-loop control module, the parsed central air conditioning fan coil unit terminal operation commands are efficiently converted into hardware action instructions, completing precise physical trigger control of heating, cooling and fan speed levels, realizing the physical action closed loop of the central air conditioning fan coil unit, and supporting the real-time adjustment needs of upper-level intelligent control.

[0206] For example, regarding the fan coil unit in the conference room on the east side of the third floor of Building A, the controller continuously monitors for a potential load surge and, in the k-th cycle, issues a high-priority pre-execution command through a predictive model, requiring the fan speed to be increased to maximum (fan speed setting value = 3), the temperature to be set to 21℃, and the cooling function to be activated. The controller-side command parsing protocol layer, based on the set command template, breaks down the operation command structure into fan speed setting commands, temperature control target commands, and cooling function bits, mapping them respectively to the PWM control channel (100% duty cycle), the temperature closed-loop adjustment port (reference voltage 0.7V), and the cooling valve output port (high level 9V). The device drive layer, according to the hardware mapping table, sequentially triggers the fan speed relay, adjusts the temperature control module, and activates the cooling valve motor, listening for presence signals. If all operations are detected to be completed within 2.5 seconds, the feedback status signal is "execution successful." The system records the action response curve, measuring a room temperature change gradient of 0.22℃ / min and a fan speed increase response delay of 1.7 seconds after the control action, meeting the intelligent control system's set targets for response speed and accuracy. Multiple field tests demonstrate that the fan coil unit controller, employing the aforementioned chain-style command mapping and hardware-level closed-loop, can stably achieve proactive control actions in scenarios of sudden load changes, with the actual physical effect highly consistent with the predicted target.

[0207] S8.5: Collect feedback data from the fan coil unit terminal after executing operation commands in real time, including action response delay, temperature control change curve and actual wind speed adjustment. Through the control command execution feedback mechanism, report the changes in operating status to the control master station to realize the closed loop of feedback data link.

[0208] Step S9: After periodic monitoring and control, the actual feedback data is compared with the predicted results, including changes in actual energy consumption, comfort levels, and anomaly occurrences. The strategy network and transfer model parameters are optimized through self-evolutionary learning to achieve a closed-loop predictive intelligent control of the central air conditioning fan coil units. Specifically, this includes:

[0209] S9.1: Collect energy consumption feedback data after periodic regulation, and use the energy consumption sensor network to obtain the actual energy consumption parameter values ​​from the terminal of each fan coil unit in real time, as the system feedback input to support subsequent regulation effect analysis.

[0210] S9.2: Collect comfort feedback data after periodic adjustment, use an indoor environmental sensor array to obtain changes in room temperature and humidity response, and use this as feedback input to characterize the actual impact of the adjustment action on user comfort.

[0211] S9.3: Collect abnormal dynamic data of equipment after periodic regulation, capture abnormal features such as fan current, valve opening degree and equipment surface temperature based on fault detection algorithm, realize real-time monitoring of equipment status, and form dynamic feedback data of equipment abnormality occurrence.

[0212] S9.4: Compare the actual energy consumption feedback parameters, comfort feedback parameters, and equipment anomaly feedback parameters with the output results of the multi-objective neural network prediction model, and use residual analysis and trend difference quantification methods to obtain the accuracy index of the model prediction and the deviation measure of the current control strategy.

[0213] S9.5: Based on the actual feedback data and the predicted residual, the self-evolutionary learning mechanism is triggered, and the adaptive gradient optimization algorithm is used to fine-tune the weight parameters of the reinforcement learning strategy network to further conform to the actual working condition changes and optimize the control decision accuracy.

[0214] S9.6: By combining the residual characteristics of operating conditions and historical learning curves, the parameter prior weights and migration strategies in the cross-domain knowledge transfer module are automatically adjusted to improve the adaptability of the transfer model to new scenarios, dynamic loads and equipment aging environments.

[0215] S9.7: Periodically refresh the optimized reinforcement learning strategy network parameters and knowledge transfer module parameters to the main control unit of the central air conditioning fan coil predictive intelligent control system to complete the closed-loop self-evolution control logic and ensure the system's adaptive evolution and continuous optimization under various operating conditions.

[0216] 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.

[0217] 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 element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its 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.

[0218] 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 intelligent control of central air conditioning fan coil units, specifically comprising: S1: Collect environmental parameter data from multiple locations within the terminal of the central air conditioning fan coil unit; S2: Perform time-series denoising, normalization, and outlier removal on the environmental parameter data to obtain a preprocessed dataset; S3: Based on the preprocessed dataset, extract load change features, comfort indicators and energy consumption trend features for different room types, area attributes and equipment aging status, and generate multi-dimensional working condition labels; S4: Input the multi-dimensional working condition labels into the cross-domain knowledge transfer model, and use the energy consumption and comfort adjustment parameters learned in the known scenario to perform knowledge transfer and parameter pre-initialization in order to obtain the parameter prior of the current target environment. S5: Based on a neural network time-series prediction model, combined with internal sensor data and external dynamic data, the load trend and equipment anomaly probability of the room in a future preset time window are predicted, and the load change risk value and anomaly tendency characteristics are output. S6: Construct a real-time digital twin simulation unit, and based on the prior parameters and the load mutation risk value, perform virtual calculations on the execution consequences of various adjustment strategies, predict the impact of fan coil unit adjustment actions on temperature control response and energy consumption, and determine the optimal adjustment timing window and activation threshold. S7: Using the optimal adjustment timing window and activation threshold as input, dynamically optimize the policy network in the reinforcement learning controller to optimize the output pre-execution control commands for the fan coil unit.

2. The intelligent control method for central air conditioning fan coil units according to claim 1, characterized in that, The environmental parameter data includes room temperature and humidity, fan current, valve opening degree, equipment surface temperature, room occupancy statistics, external meteorological data, and building orientation information.

3. The intelligent control method for central air conditioning fan coil units according to claim 1, characterized in that, Step S7 is followed by step S8: according to the pre-execution control command, an operation command is issued to the central air conditioning fan coil terminal in real time to activate the heating, cooling and fan speed adjustment commands in advance, so that the system can dynamically respond to the predicted load changes or equipment abnormalities.

4. The intelligent control method for central air conditioning fan coil units according to claim 2, characterized in that, The external meteorological data integrates weather temperature, humidity, wind speed, and air pressure.

5. The intelligent control method for central air conditioning fan coil units according to claim 1, characterized in that, Step S4 specifically includes: The multi-dimensional working condition labels generated during the multi-source information fusion process are normalized using a label standardization algorithm based on the corresponding regional attributes, room type, and equipment aging status of the labels to obtain label feature vectors suitable for cross-domain migration. Using the standardized label feature vector as input and the energy consumption characteristic parameters and comfort adjustment parameters accumulated from historical known scenarios as the source domain, a set of prior adjustment parameters that are highly correlated with the current target label is extracted. The prior adjustment parameters are retuned and adaptively weighted to eliminate potential distribution deviations between the source domain and the target environment in terms of energy consumption control and comfort indicators. Based on the obtained parameter priors, combined with the label feature vectors and room dynamic condition features, the indicators within the parameter priors are clustered and archived, and key parameters are screened. The final selected parameter prior set is input into the neural network time series prediction model and the digital twin simulation unit to initialize model parameters and simulation scenario settings, thereby achieving system-level dynamic adaptive initialization and improved cross-scenario control capabilities.

6. The intelligent control method for central air conditioning fan coil units according to claim 1, characterized in that: S1 includes using a distributed room temperature and humidity sensor network to acquire temperature and relative humidity data from multiple sampling points in the room, forming a structured temperature and humidity parameter matrix.

7. The intelligent control method for central air conditioning fan coil units according to claim 1, characterized in that: The timing window length of the digital twin simulation output is 120 seconds.

8. The intelligent control method for central air conditioning fan coil units according to claim 1, characterized in that: The S2 includes performing time-series noise suppression on the collected multi-channel environmental parameter data using an adaptive sliding filter algorithm, and performing channel-specific normalization using Z-score normalization or Min-Max normalization.

Images