Intelligent regulation and control method for fan coil of central air conditioner

Abstract

The invention relates to an intelligent regulation and control method for a fan coil of a central air conditioner, relates to an internet of things technology, and aims to solve the problems that multi-source environment parameter data is large in noise, working condition modeling is inaccurate, and equipment regulation response and energy consumption optimization are not coordinated. A regulation and control system integrating multi-parameter distributed sensing, data preprocessing, feature extraction and cross-domain knowledge migration is provided. A high-quality structured data set is formed through multi-dimensional real-time acquisition of room temperature and humidity, fan current, valve opening, human traffic, external weather and building information and in combination with denoising, normalization, alignment and missing completion. And a multi-target neural network and digital twinborn simulation are utilized to realize load trend prediction, equipment abnormity early warning and virtual deduction of an adjustment strategy, an optimal adjustment time sequence window and an activation threshold are output, and dynamic adaptive control is realized in combination with reinforcement learning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent regulation and prediction of central air conditioning systems in the Internet of Things technology, and particularly relates to an intelligent regulation method for fan coil units of central air conditioning systems. BACKGROUND

[0002] Current central air conditioning systems are widely deployed in building construction, commercial office buildings and other scenarios. Fan coil units, as terminal equipment, bear the key responsibility of adjusting indoor temperature and humidity and achieving regional micro-environment comfort. With the increasing demand for intelligent buildings and green energy-saving in the industry, intelligent optimization and regulation of central air conditioning terminals have gradually become the focus of the industry.

[0003] Some systems also explore data aggregation and distributed control strategies based on the Internet of Things platform, with basic fault diagnosis and energy-saving mode self-adaptation functions. Although there are basic automatic control and some intelligent optimization applications, the representative solutions of existing systems mainly focus on local optimization in a single scenario. On the one hand, data-driven intelligent regulation models rely on local historical data for learning, and the parameter "cold start" stage is lagging behind, making it difficult to achieve ideal performance in new scenarios, atypical rooms or equipment aging conditions. On the other hand, existing prediction models have limited ability to perceive and quickly adjust to load mutations, spatial dynamic changes and external environmental changes. The strategy update lags behind the actual energy consumption and comfort demand mutations. Current data-driven regulation strategies are often "lagging behind" demand changes, making it difficult to achieve active prediction and timely response to load mutations, extreme weather influences or sudden increases in human density. For long-term dynamics such as equipment aging and abnormal trends, the regulation system often reacts slowly, resulting in high energy consumption, large fluctuations in comfort, and poor user experience. SUMMARY

[0004] The present application provides an intelligent regulation method for fan coil units of central air conditioning systems, which aims to solve the problems of the existing technology mentioned in the background.

[0005] The present application provides an intelligent regulation method for fan coil units of central air conditioning systems, which specifically includes: S1: Collecting environmental parameter data at multiple positions in the terminal of the fan coil unit of the central air conditioning system, the environmental parameter data including room temperature and humidity, fan current, valve opening, equipment surface temperature, room traffic statistics, external meteorological data and building direction information.

[0006] S2: Time series denoising, normalization and outlier rejection processing of the environmental parameter data to obtain a data set suitable for cross-scenario modeling.

[0007] S3: Based on the pre-processed data set, a multi-source information fusion feature extraction process is constructed for different room types, regional attributes and equipment aging state, load change characteristics, comfort indicators and energy consumption trend characteristics are extracted, and multi-dimensional working condition labels are generated.

[0008] S4: The multi-dimensional working condition label is input into the cross-domain knowledge transfer model, and the energy consumption and comfort adjustment parameters learned in the known scene are used for knowledge transfer and parameter pre-initialization to obtain the parameter prior of the current target environment.

[0009] S5: Based on the neural network time series prediction model, the load trend and equipment abnormal probability of the room in the future preset time window are predicted based on internal sensor data and external dynamic data, and the load mutation risk value and abnormal tendency characteristics are output.

[0010] S6: A real-time digital twin simulation unit is constructed, and according to the parameter prior and mutation risk value, the consequences of executing various adjustment strategies are virtually calculated, the influence of fan coil adjustment action on temperature control response and energy consumption is predicted, and the best adjustment time window and activation threshold are determined.

[0011] S7: The best adjustment time window and activation threshold are input to dynamically optimize the policy network in the reinforcement learning controller, and according to the three-element adaptive adjustment mechanism of "expected energy consumption benefit-strategy execution lag-comfort tolerance", the pre-execution control instruction of the fan coil is output.

[0012] S8: According to the pre-execution control instruction, operation commands are issued in real time at the end of the central air conditioning fan coil to activate heating, refrigeration and air speed adjustment instructions in advance, so that the system dynamically responds to load mutations or equipment abnormalities in prediction.

[0013] S9: Periodically monitor the actual feedback data after regulation and control, compare the actual energy consumption change, comfort level and abnormal occurrence dynamics with the aforementioned prediction results, optimize the policy network and transfer model parameters through self-evolution learning, and realize the predictive intelligent regulation and control of central air conditioning fan coil. Closed loop.

[0014] The central air conditioning fan coil intelligent regulation and control method provided by the application has the following beneficial effects: (1) This invention achieves spatiotemporal structured perception of environmental parameters in each area of ​​the room by deploying a sensor network at multiple points at the end of the fan coil unit, combining multi-dimensional data such as room traffic, external weather, and building attributes. Combined with multi-layer preprocessing algorithms such as time series 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. Measured data show that after data processing by 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 that of traditional solutions.

[0015] (2) Through clustering and partitioning feature mapping and a multi-dimensional working condition labeling system, the present invention can automatically archive high-dimensional features for different room types, spatial attributes, and equipment health status, achieving comprehensive extraction of load changes, comfort indicators, and energy consumption trends. Utilizing cross-domain adaptive transfer learning and distribution consistency parameter mapping, the present invention can efficiently transfer energy-saving control experience and comfort adjustment weights in historical scenarios to new environments, achieving rapid cold start and initial high-quality strategic pre-response.

[0016] (3) Through the use of neural networks, a joint time-series forecast of load trends, comfort, and equipment abnormality risks can be conducted, effectively detecting sudden load changes and the probability of abnormal events 5-30 minutes in advance. The prediction error is significantly reduced compared to existing single-target prediction methods, and the system is more sensitive to changes in the macro and micro environments and sudden changes in personnel behavior.

[0017] (4) The lightweight digital twin simulation unit constructed enables virtual calculation of the response results of different adjustment strategies, automatically deducing temperature changes, energy consumption curves, and time lag effects, greatly improving the scientific and proactive nature of strategy adjustments. The adjustment timing window and activation threshold determined with the assistance of simulation results can minimize the lag in strategy execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Attachment Figure 1 The present invention is a main flow chart of an intelligent control method for a central air-conditioning fan coil unit.

[0019] Attachment Figure 2 The invention relates to a sub-flow chart of an intelligent control method for a central air-conditioning fan coil unit.

[0020] Attachment Figure 3 This is another sub-flowchart of the intelligent control method for central air-conditioning fan coil units. DETAILED DESCRIPTION

[0021] Embodiments of the present application are described in detail below with reference to several drawings. The embodiments of this application described herein are exemplary and illustrative only and are not intended to be limiting of the application. For the purpose of clarity, not all of the detailed

[0022] The disclosure hereinafter provides many different embodiments or examples for implementing different structures of the present application. For the purpose of simplification of the present application, the description of the specific examples hereinafter is described. Of course, they are only examples and the purpose is not to limit the present application. In addition, the present application can repeatedly refer to the same reference numerals and or reference letters in different examples. Such repetition is for the purpose of simplification and clarity, and it does not indicate the relationship between the various embodiments and or settings discussed. In addition, the present application provides various specific examples of processes and materials, but those of ordinary skill in the art can realize the application of other processes and or the use of other materials As shown in the accompanying Figure 1 The present application provides an intelligent control method for central air conditioning fan coil, specifically comprising: S1: Collecting environmental parameter data of multiple positions at the end of the central air conditioning fan coil, the environmental parameter data including room temperature and humidity, fan current, valve opening, equipment surface temperature, room traffic statistics, external weather data and building direction information.

[0023] S2: Time series denoising, normalization and outlier rejection processing are performed on the environmental parameter data to obtain a data set suitable for cross-scene modeling.

[0024] S3: Based on the pre-processed data set, a feature extraction process of multi-source information fusion is constructed for different room types, regional attributes and equipment aging states, and load change characteristics, comfort indicators and energy consumption trend characteristics are extracted, and multi-dimensional working condition labels are generated.

[0025] S4: The multi-dimensional working condition label is input into the cross-domain knowledge transfer model, and the energy consumption and comfort adjustment parameters learned in the known scene are used for knowledge transfer and parameter pre-initialization to obtain the parameter prior of the current target environment.

[0026] S5: The load trend and equipment abnormal probability of the room in the future preset time window are predicted based on a neural network time series prediction model combined with internal sensor data and external dynamic data, and the load mutation risk value and abnormal tendency characteristics are output.

[0027] S6: Construct a real-time digital twin simulation unit, according to the parameter prior and mutation risk value, the execution consequences of various adjustment strategies are virtually calculated, the influence of fan coil adjustment action on temperature control response and energy consumption is predicted, and the optimal adjustment timing window and activation threshold are determined.

[0028] S7: The optimal adjustment timing window and activation threshold are input to dynamically optimize the policy network in the enhanced learning controller, and the expected energy consumption benefit-strategy execution lag-comfort tolerance three-element adaptive adjustment mechanism is used to optimize the pre-execution control instructions of the fan coil.

[0029] S8: According to the pre-execution control instructions, operation commands are issued in real time to the central air conditioning fan coil end to activate heating, refrigeration and air speed adjustment instructions in advance, so that the system dynamically responds to load mutations or equipment abnormalities in the prediction.

[0030] S9: Periodically monitor the actual feedback data after regulation, compare the actual energy consumption change, comfort level and abnormal occurrence dynamics with the aforementioned prediction results, optimize the policy network and migrate model parameters through self-evolution learning, and realize the predictive intelligent regulation closed loop of central air conditioning fan coil.

[0031] The step S1: Collecting environmental parameter data at multiple positions in the central air conditioning fan coil end, the environmental parameter data including room temperature and humidity, fan current, valve opening, device surface temperature, room traffic statistics, external weather data and building direction information. Specifically, it includes: S1.1: Distributed data collection is performed on the room temperature and humidity sensor network to obtain temperature data and relative humidity data at each sampling point in the room, forming a room temperature and humidity original parameter matrix, and realizing structured perception of the spatial and temporal distribution of environmental temperature and humidity.

[0032] The input data includes the original analog or digital signals output by the distributed temperature and humidity sensor network in the room, as well as the spatial structure layout diagram of the central air conditioning fan coil end and the corresponding sampling point number.

[0033] A distributed multi-point temperature and humidity collection method is used (parameters: typical layout detection points N≥5 in the room, temperature accuracy ±0.1℃, humidity accuracy ±1.5%RH) to realize real-time acquisition of environmental parameters at different spatial positions in the room.

[0034] Further, through bus communication (such as RS485 / BACnet) or wireless networking (such as ZigBee, LoRa, etc.), the data of all distributed temperature and humidity probes are synchronously transmitted to the data acquisition master control unit, and time stamping is performed to obtain synchronous sampling multi-node original temperature and humidity data stream.

[0035] The temperature and humidity data of each sampling point are automatically collected and short-time averaged by using multi-channel time sequence buffer and periodic polling collection strategy, so as to realize local filtering and effective sampling of data redundancy.

[0036] Further, by using a spatial structure mapping algorithm, the temperature and humidity data of each sampling point are attributed to physical space coordinates and integrated into a structured room temperature and humidity original parameter matrix, as follows: wherein, represents the temperature of the i-th sampling point at the moment of t, i = 1, 2, 3, …, N; wherein N is the total number of sampling points.

[0037] represents the relative humidity value of the i-th sampling point at the moment of t, i = 1, 2, 3, …, N; wherein N is the total number of sampling points.

[0038] The spatial interpolation and trend statistical algorithm (such as Kriging interpolation method or inverse distance weighting IDW) is used to reconstruct the continuous temperature and humidity distribution field in the three-dimensional domain of the room, so as to realize the structured perception of the spatio-temporal distribution characteristics.

[0039] Through the above chain derivation, the original temperature and humidity data collected by distribution are converted into a high-dimensional, structured, time and space labeled temperature and humidity original parameter matrix, which provides support for the central air conditioning fan coil end environment monitoring and improves the perception accuracy of the system to the room environment changes.

[0040] For example, in a 15 m² standard office room, 6 temperature and humidity sampling points are arranged, the collection period is 30 s / time, and 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 the predetermined protocol, and forms the following temperature and humidity parameter matrix: The Kriging spatial interpolation method is used to map the distributed point data to the entire room area, and a spatial variation distribution surface with a temperature gradient of 0.5℃ is reconstructed, and the humidity spatial variation is not more than 2%. The sampling results show that the system can output high-precision, structured temperature and humidity distribution parameter matrix within less than 30 s delay, which provides complete and accurate data guarantee for subsequent dynamic energy consumption prediction and comfort evaluation.

[0041] ​​​​​​S1.2: Perform synchronous data acquisition operations on the fan current sensor acquisition module, obtain the fan current original parameter array by real-time acquisition of the fan motor current signal, and realize real-time quantification of the fan operating condition and energy consumption level.

[0042] S1.3: Collect feedback signals from the valve opening intelligent terminal, obtain fan coil valve opening data, form a valve opening parameter list, and accurately quantify the execution status of the terminal cold and hot medium flow regulation.

[0043] S1.4: Perform periodic temperature measurement and collection on the distributed temperature sensor array on the device surface to achieve time-series aggregation of the original data on the device surface temperature, which is used to determine the subsequent device heat dissipation status and abnormal operation trends.

[0044] S1.5: Capture data from the room's occupancy counting camera module or infrared sensor array, and use the people detection algorithm to obtain the time series of room occupancy counting parameters to characterize the room's load demand and usage scenarios.

[0045] The input data includes the digital signal output of the crowd counting camera module or infrared sensor array installed in typical traffic areas such as the ceiling of the room, doorways, and corridors, as well as the spatial layout diagram and sampling point number of the central air-conditioning fan coil terminal.

[0046] A high-precision people detection algorithm is used (parameter settings: camera frame rate ≥ 15 fps, infrared array resolution 16×16, statistical cycle 30 s / time) to realize the detection and counting of human objects in the main passages and entrance areas of the room.

[0047] Furthermore, through the convolutional neural network (CNN) or YOLOv5 target detection algorithm, combined with the region of interest segmentation (ROI Segmentation) method, the collected camera images or infrared dot array signals are analyzed in real time to identify human targets in the image / array and output the number of human targets in each frame. .

[0048] Using time series buffering and sliding average processing (sliding window length Frames), the number of human targets in consecutive frames is counted to obtain the robust human flow time series parameters: in, For the moment The corresponding smoothed statistical value of pedestrian flow. is the number of human targets detected at time (tk), K is the time offset; is the normalization coefficient, and the accumulated results are averaged so that the output is the “average number of human bodies” in these W frames.

[0049] Further, by the spatial layout mapping algorithm, the people flow statistical data of each sampling point or monitoring section is regionally archived according to the physical space division to form a room-wide people flow parameter time series vector, facilitating subsequent load demand modeling and scene recognition.

[0050] Identity deduplication and dwell time threshold elimination methods are adopted to filter repeated counting or long-stay human targets, improve the uniqueness and timeliness of the people flow statistical parameters, and achieve a more accurate characterization of the actual room load demand.

[0051] Through the above algorithm processing, the sensor or camera signal of the previous step is converted into structured people flow statistical parameter time series data, realizing high-precision dynamic representation of room usage scenarios and immediate load demand.

[0052] For example, in a 20 m² conference room, one infrared people flow array sensor (resolution 16x16, refresh rate 30 fps) is arranged at the door and one at the center aisle, and statistical results are collected every 60 seconds. Using the real-time human body detection algorithm based on YOLOv5, during a typical meeting, the peak detection of the door sensor is 15 people entering and leaving, and the peak detection of the center sensor is 13 people. Using a sliding window length frame data smoothing method, the real-time people flow statistical parameter vector is calculated In 50% of the time period during the meeting, the value is not less than 8 people, effectively reflecting the fluctuation of the room population density. Through the interface, the people flow statistical results are synchronized into the main control unit structured database, providing quantitative support for subsequent load demand prediction, comfort adjustment and energy saving strategies. The measurement shows that the error rate of the people flow parameter after identity deduplication is less than 5%, which is much better than the traditional single-point laser counting scheme. The dynamic output of this statistical parameter provides high-resolution time series data for load prediction, achieving high-precision closed-loop room usage scenario recognition.

[0053] S1.6: Call the weather data grabbing module for external weather information interface, integrate external weather parameters such as temperature, humidity, wind speed, and air pressure, realize structured input of external weather data, and provide boundary conditions for load prediction and energy consumption analysis.

[0054] S1.7: Retrieve room orientation, external wall attributes, and other building orientation parameters from the building information database, and integrate them with the above room temperature and humidity, weather data to generate a building orientation information table, which is used to improve space distribution and environmental impact factor modeling.

[0055] The step S2: performing time series denoising, normalization and outlier elimination processing on the environmental parameter data to obtain a data set suitable for cross-scene modeling. Specifically, it includes: S2.1: A multi-channel adaptive sliding filtering algorithm is used to suppress the time series noise of the collected environmental parameter time series data (including room temperature and humidity, fan current, valve opening, equipment surface temperature, pedestrian flow statistics, external meteorological data, and building orientation information). This reduces the information noise introduced by sensor fluctuations, short-term external disturbances, and sudden data anomalies, and achieves smooth output of the filtered time series results of the environmental parameters.

[0056] The collected time series data of environmental parameters including room temperature and humidity, fan current, valve opening, equipment surface temperature, room traffic statistics, external meteorological data and building orientation information serve as input for processing in this step.

[0057] Adopting multi-channel adaptive sliding filtering 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), it realizes noise suppression and signal smoothing processing of various types of acquisition channel data. The algorithm sets the window length W independently in each acquisition channel. k , by counting the historical sampling points in the current window, the impact of spikes and instantaneous interference on time series data is reduced.

[0058] Furthermore, by analyzing the original time series data x detected by each channel k (t) Adopt adaptive weighted sliding average filtering to achieve automatic adjustment of distortion peak and smooth data output. Specifically, for each data channel k, the following processing is adopted: in is the time series data processed by adaptive weighted moving average filtering, is the kth channel moment The original sampling data, It is an adaptive weight coefficient, and the weight is dynamically adjusted according to the data fluctuation degree, autocorrelation and pre-set channel characteristics in the window.

[0059] Furthermore, through the outlier sensitivity suppression algorithm, the weight of the sudden change sample points (such as outliers with jumps greater than 3 times the mean square deviation) is adaptively reduced during the filtering process to avoid abnormal sudden disturbances on the smooth output. In the continuous sampling period, the historical noise distribution modeling method is used to calculate the historical fluctuation standard deviation σ of the acquisition channel. k , dynamically update the sliding window weight distribution so that the filtering result takes into account both sensitivity and robustness.

[0060] Further, periodic consistency check is performed on the multi-channel filtered time series results at the global level. Redundant channel cross-checking is used to analyze the consistency of parameter groups with strong spatial correlation (such as temperature and humidity at multiple points, fan current at different positions), and local encryption smoothing is triggered immediately when abnormal distribution is found to achieve distribution optimization under sudden interference.

[0061] Through the above multi-step chain derivation, the original collected multi-channel environmental parameter time series data is converted into high-quality time series parameters after noise suppression, sudden abnormal buffering and smoothing output, providing a stable data foundation for subsequent normalization, consistency and abnormal point removal processing, and realizing data quality assurance for cross-scene modeling.

[0062] For example, in a 20 m² standard conference room environment, 6 sensing channels (temperature, humidity, fan current, valve opening, equipment surface temperature, and people flow statistics) are configured, and the collection interval is 30 seconds. For the temperature and humidity channels, the sliding window length is set to W=7, and the initial weight is equally divided. At a certain time, the temperature data sequence is [23.6, 23.7, 23.8, 23.7, 26.5, 23.7, 23.8], and the 5th value is affected by sensor transient abnormality, which is much higher than the average. After setting the 3σ threshold, 23.7±3×0.1, the abnormal value weight is automatically reduced to 0.1, and the weights of other normal points are each 1. After weighted moving average, the filtered output is 23.74°C, effectively suppressing the sudden abnormal point. The fan current channel uses W=5, the historical standard deviation is 0.02A, and the global weight factor is reduced to 0.5 during the sudden abnormal period, achieving smooth output of data during the fan starting operation. After processing of all channels, the noise variance of the original data is reduced by 60%, and the spot error rate of the people flow channel is reduced to less than 3%. Through this step, the multi-channel original environmental parameters are output as smooth, abnormal-free high-confidence time series results, providing a solid foundation for subsequent normalization and abnormal point detection.

[0063] S2.2: Based on professional normalization operators (such as Z-score standardization or Min-Max normalization), the sliding filtered environmental parameter time series results are mapped to a unified dimension interval for consistency scale mapping, eliminating the scale inconsistency caused by device models, layout differences, and collection hardware output range, and obtaining a normalized environmental parameter matrix.

[0064] S2.3: Based on joint confidence interval detection and moving window statistical adaptive threshold algorithm, the normalized environmental parameter matrix is processed for time series data abnormal point removal, identifying and removing abnormal sample points caused by collection device failure, outlier peaks, and extreme external impact, generating a high-confidence environmental parameter available data set without abnormal points, to ensure the rationality of data distribution and the robustness of subsequent feature extraction.

[0065] On the basis of the normalized environment parameter matrix, a joint confidence interval detection method (parameter setting: confidence level a = 0.95) is used to extract the statistical characteristics (such as mean μ k , standard deviation s k ) of the time series data of each channel, and the upper and lower confidence limits of each channel in the current sliding time window W n are calculated to form a dynamic detection threshold interval.

[0066] Further, through the moving window statistical adaptive threshold algorithm (sliding window length L w = 50~100, adjusted according to the data sampling frequency), the real-time window statistics of each time series data point x k (t) are obtained, and the corresponding expected range is obtained: [ , ] Wherein, and are the mean and standard deviation of the kth channel in the sliding window W n , is the quantile coefficient under normal distribution.

[0067] Using a joint decision mechanism, data points in the normalized matrix n k (t) that fall outside the dynamic confidence interval are marked as abnormal candidate points, and abnormal candidate points of valve opening, fan current and other multi-channel data are cross-compared to aggregate device categories, spatial regions, data categories and other collaborative abnormal features to enhance the robustness of abnormality discrimination.

[0068] Further, for abnormal candidate points, a continuous peak value recognition and extreme external impact detection method is used to eliminate extreme samples caused by short-term disturbances (such as sudden power failure, equipment switching transient), and combined with intraday periodic pattern analysis, to prevent normal periodic fluctuations from being misjudged as abnormal.

[0069] According to the abnormal tolerance configured by the collection device administrator (for example, the allowed proportion of extreme fluctuations is <0.5%), after multiple screenings, data distortion points caused by collection hardware failure, long-time communication disconnection, strong sensor interference, etc. are completely removed, and the remaining high-confidence environment parameter data points are merged into an environment parameter available data set without abnormal points.

[0070] Through the above joint confidence interval detection and adaptive moving window abnormal point removal algorithm, the normalized data matrix n k (t) of the previous step is converted into a high-confidence, abnormal-point-free data set, realizing the rationality of the environment data distribution and the robustness of the subsequent feature extraction.

[0071] Exemplarily, in a certain 25 m² office, 5 channels of temperature and humidity, fan current, valve opening, equipment surface temperature and human flow are configured, the sampling period is 30 s, and the normalized matrix n is generated by long-time running and accumulation k (t), totaling 2880 points / day. The confidence interval method is set at a = 0.95, for the room temperature channel, the daily average μ T = 23.65, the standard deviation σ T = 0.46, and the dynamic confidence interval range is [22.75, 24.55] through a sliding window (L w = 100) statistics. A total of 14 outliers are detected in the actual historical data, distributed in the morning and evening peak of equipment switching and the sensor restart period. For the fan current channel, 5 obviously abnormal peak points are further removed through equipment type cross comparison and peak identification. Finally, the proportion of abnormal points of all channels is <0.5%, and 2870 high-confidence environmental parameter data are output, meeting the robust data input requirements of subsequent multivariate feature extraction and load prediction models. The application results show that after the removal of abnormal points, the subsequent modeling MSE is reduced by about 22%, the model generalization error converges significantly faster, and the system's perception ability of weak features and mutation risks is significantly improved.

[0072] S2.4: Based on the high-confidence environmental parameter available data set without abnormal points, a multi-source hierarchical missing value completion algorithm (such as K-neighbor interpolation and historical working condition playback alternately selected) is used to fill in the missing intervals in individual time periods caused by short-term communication interruption or sensor failure, and a full-dimensional completed environmental parameter time series set is output to ensure that the downstream modeling task obtains continuous and complete information input.

[0073] The missing value completion processing is performed on the high-confidence environmental parameter available data set without abnormal points to cope with the situation of incomplete time series data caused by short-term failure of fan coil sensors, communication interruption and other factors.

[0074] A multi-source hierarchical missing value completion algorithm (parameters: priority sequence K-neighbor interpolation method K = 5, historical working condition playback window W r = 24h) is used to realize the preliminary estimation and filling of the missing intervals detected in all collection channels (such as temperature and humidity, fan current, valve opening, etc.).

[0075] Further, through the K-neighbor interpolation method, the same type of data of the nearest K non-missing time points are taken as the reference, the sample similarity is scored according to the Euclidean distance measurement or cosine similarity, and the data value of the missing time point is estimated by weighted average: wherein, is the missing data point to be filled, for the selected first K reference non-missing samples, represents a normalized weight determined by distance or similarity.

[0076] Further, for long time periods or missing intervals across sampling periods, a historical working condition playback interpolation method is adopted to use the past W r The stored working condition sequence corresponding to the time period of the window is synchronously aligned. By collecting the time sequence fragments under the same time and same parameters in the historical working condition, mean playback filling is performed to enhance the time sequence continuity and physical rationality of the filling.

[0077] Further, for the case of multi-channel synchronous missing, through a hierarchical priority mechanism, the core acquisition channels (such as temperature and humidity, fan current channels) are filled first, and then the auxiliary channels (such as passenger flow, weather, building direction) are filled, to ensure the integrity and trend coherence of the main control parameter sequence.

[0078] Further, for all newly generated or completed data intervals, a consistency constraint checking strategy is adopted, such as sliding window smoothness test, time sequence jump detection and physical upper and lower bound review, to screen whether the interpolation result causes sequence discontinuity, out-of-limit jump or physical unreasonable situation, and to feedback secondary correction for abnormal situation until it meets the global data distribution and system operation physical properties.

[0079] Through the multi-source hierarchical missing value completion algorithm, the data set without exception but with missing data is processed into a complete time sequence set of environmental parameters with full dimensions, which meets the needs of continuous, full-quantity and high-confidence data for subsequent modeling tasks, and provides a solid information foundation for cross-scene modeling, feature engineering and strategy optimization.

[0080] For example, in a 700 m^2 open office, the collection frequency of temperature and humidity, fan current and valve opening is every 30 seconds, the passenger flow statistics and weather data collection interval is every 60 seconds. In the actual deployment process, due to network switching, single 2-minute missing data appears in the temperature and humidity channel, single 5-minute missing data appears in the fan current channel, and 1-minute missing data scattered in the night appears in the valve opening channel, with a total missing rate of about 0.5%. For single-point short-time missing within 30 seconds, K=5 is taken as the K nearest neighbor interpolation algorithm, and the effective data within 2.5 minutes before and after is weighted to fill in, to ensure that the mean square error of the interpolated sequence is less than 0.03. For intervals of 5 minutes or more, historical working condition sequences under the past 24 h energy-saving mode are referred to for playback interpolation, to ensure that the filling interval change trend is fitted with the actual working day curve with R^2>0.92. All filling intervals pass the sliding window consistency test (window L=7) without obvious mutation or discontinuous jump points. The final output is a complete time sequence matrix with consistent length, full content and physical constraint verification, which saves about 30% of the abnormal screening work hours for downstream feature extraction and prediction modeling, and the MAE of the system energy consumption prediction model decreases by 4.5%.

[0081] S2.5: For the full-dimension completed environment parameter complete time sequence set, use time synchronization alignment algorithm (such as multi-channel dynamic time warping, DTW) and master clock multi-source alignment technology to perform time sequence data alignment of multi-channel acquisition channels to unify time labels and sampling steps, realize synchronization and cooperation of different source parameters, and output high-quality structured data set with time sequence consistency guarantee as the standard input for subsequent cross-scene modeling and knowledge transfer.

[0082] The step S3: based on the pre-processed data set, for different room types, regional attributes and equipment aging states, a feature extraction process of multi-source information fusion is constructed, load change characteristics, comfort indicators and energy consumption trend characteristics are extracted, and multi-dimensional working condition labels are generated. As shown in Figure 2 Specifically, it includes: S3.1: For the data set that has completed preprocessing, a multi-source information grouping index is established according to the room type, regional attribute and equipment aging state, and a feature hierarchical mapping algorithm is used to archive the environment parameter data according to the grouping to obtain data partitions subdivided according to application scenarios, laying a prior data foundation for subsequent working condition similarity matching and specific strategy optimization.

[0083] For the full-dimension completed environment parameter complete time sequence set, a multi-source information grouping index is established with the room type, regional attribute and equipment aging state as the grouping variable to realize structured archiving of the data source.

[0084] A feature hierarchical mapping algorithm (parameters: grouping dimensions cover room function type, space partition number, sensing device identifier, and equipment health degree interval) is used to classify the environment parameter data step by step, and the data in the same type of room, similar area or health degree interval is aggregated and archived.

[0085] Further, a semantic label mapping mechanism is used to correspond the above grouping data with a preset application scenario label system (such as "conference room-west zone-aging medium") one by one to generate data partitions with scene side labels.

[0086] Through a data integrity verification algorithm, the key parameters (temperature, humidity, fan current, etc.) in each data partition are checked for coverage, and the partitions with collection quantity below the set threshold are filtered out to ensure that each application scenario partition has sufficient available data quantity.

[0087] Further, a partition feature frequency analysis is used to describe the statistical distribution of load characteristic parameters, comfort index distribution and energy consumption trend distribution of each data partition, and extract statistical characteristic indicators of each partition to provide data support for subsequent working condition similarity retrieval and individualized strategy optimization.

[0088] Through the feature hierarchical mapping algorithm, the full environmental parameter data is structured into fine-grained data partitions with grouping labels, statistical characteristics and physical consistency requirements, realizing precise feature hierarchical for different application scenarios, and laying a high-confidence prior data foundation for subsequent multivariate feature extraction and scenario modeling tasks.

[0089] For example, in the air conditioning management system of a Grade-A office building, the environmental parameter data of 600 independent rooms is synchronously collected, involving three typical room types (standard office, conference room, computer room), each type of room is divided into four spatial regions (east, west, south, north), and the fan coil equipment in each room is divided into three health degree levels according to the actual service life: "brand new (0-3 years)", "slightly used (3-7 years)", and "aging (>7 years)". For the full data collected, the feature hierarchical mapping algorithm is used. First, a type-level index is constructed according to the room type field, then it is further subdivided according to the spatial region number, and finally the fan coil equipment health degree interval is introduced as a three-level index variable to merge the environmental parameters of all rooms into 36 independent partitions (3 room types x 4 regions x 3 health degrees). In the conference room-west-aging partition, the data volume is 19820, covering 6 parameters such as complete temperature and humidity, fan current, with a coverage rate of 99.2%. Through the data integrity verification algorithm, 1 partition with abnormal data volume is excluded. The feature frequency statistics of the data in the partition show that the temperature fluctuation mean square deviation is 0.43℃, the fan current average is 0.22A, and the equipment surface temperature distribution deviation is high, reflecting that the room operation and maintenance load is high and the aging signs are obvious. Finally, the data in each partition has the characteristics of scenario scripting application, and outputs multi-source information partitions indexed by room type-region-health degree three-structure labels, providing high-confidence grouping input data for subsequent special modeling and global transfer learning strategies, which can support the system performance improvement of reducing artificial inspection mean value by 30% and improving aging interval risk warning accuracy by 15%.

[0090] S3.2: Using the multi-source information partition environmental parameters as input, a multivariate dynamic load modeling algorithm is used to extract the key load variation characteristics of the room, including cold / heat load time series variation, fan coil power fluctuation, and valve response amplitude, etc. The output is a load variation characteristic parameter sequence, which provides high-dimensional feature input for energy scheduling and load risk prediction.

[0091] S3.3: Based on the same multi-source information partition environmental parameters, the indoor comfort evaluation index system (such as PMV, PPD index, etc. human comfort algorithm) is used to calculate comfort indicators such as temperature, humidity, wind speed, and surface temperature difference, and generate an indoor comfort feature parameter table for each time period, providing input boundaries for the target weight of the control strategy.

[0092] S3.4: Based on the aforementioned load variation characteristics and comfort index characteristics, a time-series energy consumption decomposition algorithm and sliding window statistical analysis are used to extract features of energy consumption-related parameters such as fan current and valve opening, generate energy consumption trend feature vectors, and provide core constraint data for subsequent energy consumption prediction and energy efficiency tuning models.

[0093] S3.5: Using the obtained load change characteristic parameters, comfort characteristic parameters, and energy consumption trend characteristic vectors as input, the data is mapped to generate multidimensional operating condition labels through multidimensional operating condition label generation rules. The label content includes room partition number, current equipment health, load type, comfort level, energy consumption level, and other information, providing a structured scenario description for cross-domain knowledge transfer and strategy optimization models.

[0094] Taking 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 is adopted (parameters: label dimensions are agreed to be partition number, health level, load type, comfort level, and energy consumption level) to realize the structured working condition label mapping function.

[0095] Furthermore, by associating the label mapping table, we link the load change feature sequence with the comfort index and energy consumption trend value at the scene level according to the spatial number of the data partition, the equipment aging status and the partition function type. Using label encoding algorithms (such as independent hot encoding and multi-level hierarchical mapping technology), we encode the spatial partition number P, equipment health H, load type L, comfort level S, energy consumption level E and other factors dimensionally to generate a preliminary label matrix T: Where N is the number of samples, and D is the label dimension. P is taken from the room partition code 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 cooling, heating, and mixed load types, S is obtained by mapping the PMV and PPD thresholds to obtain the comfort level range, and E is quantified to standard tiers using energy consumption trend characteristics.

[0096] Furthermore, through the label consistency verification algorithm (such as redundant label deduplication, confidence judgment, and label conflict decomposition), the preliminary label matrix T is deduplicated and normalized to eliminate abnormal conflicts or fuzzy duplicate labels, ensuring that each data instance corresponds to a unique and multi-dimensionally consistent label combination.

[0097] Furthermore, label synthesis and multidimensional normalization techniques are used to compress the multidimensional label encoding into a standardized label vector T*, facilitating subsequent embedded processing of knowledge transfer and policy optimization models. This label vector, in the form of sparse hot encoding or vectorized embedding, provides a complete scenario description capability, including partition location, health state differentiation, load characteristics, comfort goals, and energy consumption targets.

[0098] By the multi-dimensional working condition label generation rule, the original information of the load change characteristic parameter, the comfort characteristic parameter and the energy consumption trend characteristic vector is converted into a multi-dimensional working condition label structure, accurate and migratable expression of a complex room-equipment-energy consumption scene is realized, and high-confidence structured input data is provided for cross-domain knowledge migration and strategy optimization of the enhanced learning model.

[0099] For example, in a typical office building, 19820 scene samples of a conference room-western district-aging equipment partition are taken as objects, a load change characteristic sequence (cold load mean 3.2kW, fluctuation variance 0.7kW^2), a comfort parameter table (PMV=0.4, PPD=8%), and an energy consumption trend characteristic vector (daily average maximum power 0.78kW, energy consumption growth rate 1.5% / h) are input.

[0100] By using the label encoding algorithm, the partition number P is specified as

[03] , the health degree H is evaluated as [aging (2)], the load type L is clustered as [cold load (1)], the comfort level S is quantified as [relatively comfortable (2)], and the energy consumption level E is graded as [high (3)]. The label matrix T is: [03, 2, 1, 2, 3]. No redundancy or conflict is detected in the label redundancy detection, and the five-dimensional standardized label vector T* is directly obtained by label vector normalization processing: [0.03, 0.66, 0.33, 0.66, 1.00].

[0101] Finally, the multi-dimensional working condition label T* output by the embodiment is input into a subsequent migration learning model and a strategy network module, supports knowledge migration and energy efficiency priority regulation mechanism of the system for an aging equipment-high load-high energy consumption-local comfort extreme scene, and the accuracy of automatically generated measured labels reaches 99.8%, effectively improving the global scene recognition and parameter migration generalization performance.

[0102] The step S4: inputting the multi-dimensional working condition label into a cross-domain knowledge migration model, using energy consumption and comfort adjustment parameters learned in a known scene, performing knowledge migration and parameter pre-initialization to obtain parameters priori of a current target environment. As shown in Figure 3 The specific steps include: S4.1: based on the region attribute, room type and device aging state corresponding to the label, the label feature vector is normalized by using a label standardization algorithm to obtain a label feature vector suitable for cross-domain migration, and standardization preparation is performed for subsequent knowledge migration model input.

[0103] The cross-domain knowledge transfer model includes a domain self-adaptive model, a knowledge distillation model, a feature alignment model, a parameter sharing model, a federated learning model, a federated cross-domain recommendation model, a transfer reinforcement learning model, a relationship transfer model, and the like.

[0104] Based on the multi-dimensional working condition label generated by the multi-source information fusion module, the input conditions include the structured label matrix T* processed by the previous stage (feature hierarchical mapping, label encoding and consistency test), each label instance corresponds to a unique room partition number, device health degree, load type, comfort level and energy consumption level code.

[0105] The label standardization algorithm is adopted, and for each dimension code (including space number, health degree, load type, comfort level, and energy consumption level) in the label matrix T*, according to the information distribution characteristics and the consistency requirements of the target cross-domain scene, the normalization mapping interval (parameter: standard normalization interval [0, 1]) is configured. For numerical labels (such as device health degree and energy consumption level), Min-Max normalization processing is adopted, wherein, is the normalized specific label element, is the specific label element to be normalized, , is the minimum and maximum value of the label parameter in the whole data.

[0106] For label dimensions with discrete attributes (such as room partition, load type, and comfort level), a mixed strategy of one-hot encoding and normalized hot encoding is adopted, wherein, is the one-hot encoding value of the i-th sample in the j-th label dimension, is the corresponding normalized one-hot encoding value, is the total number of unique values of the label dimension, and the normalized encoding vector is output.

[0107] Further, through label distribution visualization and statistical consistency test, the label feature vector distribution output by batch normalization is subjected to Kolmogorov-Smirnov test or Kullback-Leibler divergence analysis, to ensure that the label normalization results under different room types, regional attributes and aging states have cross-domain scene distribution consistency, and to provide a standard template for subsequent cross-domain transfer learning model input.

[0108] By the above label standardization algorithm processing, the multi-dimensional working condition label batch is converted into a standardized label feature vector, realizing the unified scale mapping and structured expression between multi-source labels, and preparing the optimal feature basis for the input of the cross-domain knowledge transfer model.

[0109] For example, for the aforementioned office building conference room-west zone-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 degree H is normalized to 0.66, the load type L is normalized to 0.33, the comfort level S is normalized to 0.66, and the energy consumption level E is normalized to 1.00. L2 norm normalization is performed on the label vector, and the output standard label feature vector T * =[0.021, 0.464, 0.232, 0.464, 0.704]. After the normalization processing of 19820 label vectors, the distribution consistency test result Kolmogorov-Smirnov distance is less than 0.07, which meets the cross-domain input uniformity constraint. Finally, the standardized label feature vector output by this step is input into the cross-domain transfer learning model, supporting knowledge parameter mapping, realizing efficient transfer of energy efficiency strategy and comfort target under different room and equipment health status, supporting convergence speed improvement of more than 20% in actual scene strategy optimization, and label conversion accuracy reaching 99.8% in energy saving application test scene.

[0110] S4.2: Taking the standardized label feature vector as input, taking the energy consumption characteristics parameters and comfort adjustment parameters accumulated in the historical known scene as the source domain, using the domain adaptive transfer learning algorithm, comparing the implicit correlation features between different labels in the transfer learning framework, extracting the prior adjustment parameter set with high correlation with the current target label through parameter mapping and distribution matching.

[0111] Taking the standardized label feature vector T output by the label standardization algorithm as the input object, inputting it into the cross-domain knowledge transfer learning model to realize efficient transfer and matching of source domain and target domain parameters. Using the domain adaptive transfer learning algorithm (parameters: feature space consistency mapping, distribution matching loss threshold, feature extraction main core), realizing the joint alignment of energy consumption characteristics parameters and comfort adjustment parameters in the source domain (historical known scene) and the current target environment label feature space, through hidden layer feature embedding, spatial projection and principal component reconstruction of the label feature vector T and the source domain parameter distribution, generating parameter mapping feature tensor.

[0112] Further, by using the maximum mean difference (MMD) distribution matching algorithm (parameters: kernel function type selected as Gaussian kernel, feature dimension adjustment coefficient β, sample batch normalization), the source domain parameter distribution P s (X) and the target label feature domain distribution P t(X) performing adaptive distribution convergence calculation, optimizing loss function wherein, is the loss function value, is the kernel space mapping function, is the source domain kernel space mapping function, is the target label feature domain kernel space mapping function, X s is the sample feature sampled from the source domain, X t is the sample feature sampled from the label feature domain, is the expected value operator, by reducing , the distribution alignment of the source domain and the target label feature domain is realized.

[0113] Further, the parameter level mapping and feature correlation screening algorithm (parameters: label similarity score threshold η, feature interaction weight λ) is adopted, in the source domain experience parameter library (covering energy consumption curve response coefficient, comfort target control weight, dynamic threshold setting), the historical parameters are screened one by one according to the feature correlation alignment value, the historical parameters are completed the optimal projection in the label space through the mapping matrix W, and the candidate prior adjustment parameter set with high correlation is formed . The formula is: wherein, is the source domain parameter set, is the similarity measure function (cosine similarity, Euclidean distance can be applied) of the label vector and the parameter vector, is the mapping matrix, which projects the parameters from the original space to the label space or the feature space, so that the parameters and the labels can be compared, is a single parameter vector, is the projection vector of the parameter in the label space; is the target label vector; is the parameter subset selected from the source domain parameter set to maximize the similarity.

[0114] Further, through the implicit feature residual screening algorithm (parameters: error allowed interval δ, feature contribution degree ranking), the candidate parameter set obtained by mapping is subjected to residual zero processing, the historical parameter components that cannot adapt to the high weight features of the target working condition are removed, and only the prior parameter subset highly coupled with the current label feature vector T is retained . Through the above processing chain, the standardized label feature vector driven domain adaptive transfer learning algorithm, distribution matching and parameter screening technology are organically combined to obtain the prior adjustment parameter set adapted to the current target label, and the system improvement of parameter adaptation accuracy and convergence speed in the knowledge transfer process is realized. For example, in a conference room of a Grade A office building in the west area of an aging equipment environment, the input standardized label feature vector T = [0.021, 0.464, 0.232, 0.464, 0.704], reference K nearest neighbor distance weighted strategy, select the energy consumption curve response parameter and comfort regulation parameter with Euclidean distance less than 0.12 in the source domain historical scene to form the preliminary source domain parameter subset (total 67 groups). Through the maximum mean difference (MMD) distribution matching, the Gaussian kernel bandwidth parameter σ = 0.5 is selected, and after batch normalization, the source domain and target domain MMD loss can be reduced to 0.008. The correlation screening parameter η is set to 0.85, and the parameters with cosine similarity higher than 0.85 are screened out as candidate , a total of 41 groups. Based on feature residual screening, the low coupling parameter group with δ > 0.05 is removed, and finally 28 groups of highly correlated prior adjustment parameters are output. In actual system deployment, the pre-response ability of the cold start strategy in the transfer learning stage is improved by 19.2%, and the 12-hour energy efficiency optimization convergence speed is improved by 23%, and the prior parameters can effectively support the risk suppression and energy efficiency improvement goals in the aging equipment-high load-high energy consumption interval.

[0115] S4.3: Based on the prior adjustment parameter set and the target label feature vector obtained above, a transfer regularization constraint mechanism is adopted to adjust the parameters and adaptively adjust the weights of the prior adjustment parameters, eliminate the potential distribution deviation of the source domain and the target environment in energy consumption regulation and comfort indicators, and output the parameter prior adapted to the current target environment.

[0116] S4.4: For the obtained parameter prior, combined with the label feature vector and the room dynamic working condition characteristics, parameter clustering and sensitivity analysis algorithm is adopted to cluster and archive each index (such as cold load regulation sensitivity, wind speed efficiency threshold, prediction timing window) in the parameter prior and key parameter screening, to obtain a more universal and generalizable parameter prior set.

[0117] S4.5: The parameter prior set obtained by the final screening is input into the subsequent neural network timing prediction module and digital twin simulation unit, which is used to initialize the model parameters and simulation scene setting, realizes the dynamic adaptive initialization and cross-scene regulation ability improvement of the system level, and provides high reliable basic data for the optimization strategy of the reinforcement learning controller.

[0118] The step S5: the load trend and the equipment abnormal probability of the room in the future preset time window are predicted based on a neural network time series prediction model combined with internal sensing data and external dynamic data, and a load mutation risk value and an abnormal tendency feature are output. Specifically, it comprises the following steps: S5.1: The internal sensing data (including room temperature and humidity sensing data, fan current signal, valve opening signal, and equipment surface temperature signal) and the external dynamic data (including room traffic statistical data, meteorological data, and building direction information) are synchronously arranged, time series resampling and alignment algorithms are used to generate a multi-source feature time series matrix under a unified time reference, and consistent data is ensured for use in the time series prediction model.

[0119] The internal sensing data (including room temperature and humidity sensing data, fan current signal, valve opening signal, and equipment surface temperature signal) and the external dynamic data (including room traffic statistical data, meteorological data, and building direction information) are set with a unified time reference and sampling frequency to form multi-source original data input.

[0120] A time stamp alignment method (parameters: master clock reference, NTP or PTP synchronization protocol) is used to realize time synchronization of original data streams of different sensor nodes, map various data to a standard time axis, and eliminate cross-device sampling delay and clock drift.

[0121] Further, a time series resampling algorithm (parameters: target sampling period Δt, resampling interval selection linear interpolation or spline interpolation) is used to perform resampling processing on individual channels with inconsistent frequency or missing data frames, fill in all data points on the time axis, unify the data sampling frequency, and obtain equal-interval time series data sequences of each channel.

[0122] Further, a multi-source data path aggregation algorithm (parameters: multi-channel identifier, spatial position encoding, and data validity mask) is used to aggregate various time series data after processing, construct a multi-dimensional feature column with a unified channel number, eliminate invalid or abnormal sampling frames, and collect a synchronized data set of all features at time t.

[0123] An index mapping and hash archiving method (parameters: sample step, and data window length) is used to convert the multi-source aggregated data sequence into a fragmented structure according to the specified step, form a standardized time series data block containing all feature columns, and facilitate subsequent windowed feature extraction and convolution operations.

[0124] Through the above synchronization, resampling, and aggregation processing, a multi-source feature time series matrix under a unified time reference is output where N is the number of sampling steps, d is the comprehensive feature dimension, consistent input, collaborative modeling, and efficient integration of cross-source features in the multi-target time series prediction model are realized.

[0125] Exemplary, in the office building west conference room scenario, the room temperature and humidity sensor sampling period is 2 seconds, the fan current / valve opening / surface temperature sampling period is 5 seconds, the people flow camera data is 10 seconds, and the weather and building direction information update period is 60 seconds. All nodes correct the master clock through the NTP protocol, and the sampling step is 2 seconds. The fan and valve values are linearly interpolated to fill the step, and the weather and direction parameters are filled with the nearest neighbor to align the short period. Through hash channel aggregation, the data buffer outputs 30 frames of structured data every 60 second window, and each frame includes 14-dimensional channel features (temperature and humidity 4, fan 2, valve 2, surface temperature 2, people flow 1, weather 2, and direction 1). In the fault simulation scenario, the maximum sampling delay per cycle is less than 0.7 seconds, and the residual clock drift error is less than 0.001%. The final output The time sequence matrix is , and the 100% step alignment has no missing data, realizes consistent and accurate synchronization of all internal and external data, and supports the windowed dynamic modeling of the multi-target neural network backend.

[0126] S5.2: Based on the multi-source feature time sequence matrix, a feature engineering algorithm (such as sliding window statistics, change point detection, and autocorrelation analysis) is used to dynamically extract features from the time sequence signal to obtain the load change characteristics, comfort index, and energy consumption trend characteristics of the fan coil end, to form a multi-parameter input tensor suitable for further time sequence modeling.

[0127] S5.3: The multi-parameter input tensor is input into the neural network time sequence prediction model, and the long short-term memory network (LSTM) and attention mechanism are used in combination to encode the load trend and device abnormal risk in parallel, capturing the time-dependent relationship of load cascade mutation patterns and device health hazards.

[0128] S5.4: Perform multi-objective loss function optimization on the time sequence prediction results, and according to the energy consumption prediction error, comfort evaluation index error, and device abnormal risk detection accuracy, jointly optimize the weights for iteration to output the load mutation risk value and abnormal tendency characteristics in the future preset time window.

[0129] S5.5: Dynamically aggregate and attribute analyze the load mutation risk value and abnormal tendency characteristics, use mutation threshold detection and abnormal probability attribution algorithm to identify high-risk moments and corresponding dominant features, and provide decision suggestions for priority allocation of digital twin simulation and reinforcement learning strategy.

[0130] The step S6: Constructing a real-time digital twin simulation unit, according to the parameter priori and mutation risk value, virtually calculating the consequences of the execution of various adjustment strategies, predicting the influence of fan coil adjustment action on temperature control response and energy consumption, and determining the optimal adjustment time window and activation threshold. Specifically, it includes: S6.1: Initialize the operating conditions of 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 terminal.

[0131] S6.2: Utilize the synergistic effect of physical modeling methods and data-driven modeling methods to perform multi-dimensional modeling of fan coil unit adjustment strategy variables. By establishing a digital twin simulation network that covers the thermodynamic characteristics of the main unit and terminal, ventilation efficiency, and dynamic load response parameters, the fan coil unit terminal response behavior under multiple working conditions can be modeled, providing a reliable model foundation for virtual calculations.

[0132] S6.3: Based on the initialization parameters of the operating conditions and the input of the adjustment strategy variables, batch virtual calculations are performed on the execution consequences of different fan coil unit adjustment strategies (including heating, cooling, wind speed adjustment, etc.). Using real-time numerical simulation and state feedback mechanism, dynamic prediction data on room temperature changes, energy consumption changes, and indoor comfort indicators for each strategy are output.

[0133] S6.4: Perform a multi-objective time series evaluation on the temperature response curve, delay threshold, and energy consumption curve output by the virtual calculation of the strategy action. Use the working condition discrimination algorithm to standardize and quantify the benefit-cost function of the consequences of each adjustment strategy, and select the adjustment time series window that satisfies the balance between energy efficiency, hysteresis tolerance, and comfort, providing a decision reference for the subsequent determination of the strategy activation threshold.

[0134] 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 terminal adjustment strategy. This algorithm optimizes and determines the set of control actions that can be triggered in advance, and generates the optimal adjustment timing window and activation threshold parameter set that can be dynamically called by the reinforcement learning controller, achieving efficient strategy linkage with the physical system.

[0135] The input is the core operating condition evaluation results such as temperature response curve, delay threshold, energy consumption curve, etc. obtained through multi-objective timing evaluation and digital twin simulation, as well as simulation output data containing the initial parameters of the actual operating conditions and strategy variables.

[0136] A piecewise dynamic programming decision optimization algorithm (parameters: adjustment time interval, strategy action set, benefit-cost function) is used to gradually analyze the optimality of the adjustment strategy effects at different times, realize multi-objective adaptive window search, and establish the optimal pairing relationship between strategy actions and actual working condition responses.

[0137] Further, the algorithm (parameters: expected energy reduction ΔE, target comfort score ΔC, control time lag acceptance threshold ΔT) is quantified by a multi-objective benefit-cost function, and each set of regulation time window and action trigger point is standardized and quantified. The following comprehensive performance evaluation index is used: wherein, is the comprehensive performance score of the action at time , is the energy consumption change brought by the action, is the comfort improvement amount, is the response lag introduced by the control action, , , , is the weight of each target, , , is the normalized reference value.

[0138] Further, through the sliding window sensitivity analysis algorithm (parameters: window length L, time lag step δt), the local extreme section is searched in all feasible time sequence schemes, and the best regulation time window [t s , t e ] is determined and the corresponding activation threshold is solved, which satisfies the following formula: Through the joint optimality search of the sliding window and the action space, the control action set with the optimal energy efficiency and acceptable comfort under the target working condition is selected , is the starting time of the optimal time window, is the ending time of the optimal time window, is the optimal activation threshold; is the maximum operator, which means selecting the group that makes the maximum from all possible time window and threshold combinations , is the starting time of the time window, is the ending time of the time window, is the activation threshold; is the target performance function, which comprehensively evaluates the weighted performance of energy efficiency and comfort when the time is t and the threshold is ; is the constraint condition, which means that the performance of all time points in the entire time window must meet certain requirements (such as comfort not lower than a certain lower limit), rather than only optimizing a certain moment.

[0139] Further, the optimal adjustment timing window and activation threshold are outputted in a multi-layer structure according to rooms, partitions and policy types by adopting an action parallel grouping and archiving mechanism (parameters: room ID, policy label, priority layering), to generate a parameter set that can be directly and dynamically called by the enhanced learning controller.

[0140] Through the above decision optimization algorithm and timing sensitivity analysis method, the timing evaluation and simulation output results are converted into the optimal adjustment timing window and activation threshold parameters, to realize the early triggering of the fan coil terminal control action and the intelligent linkage of the policy, and significantly improve the ability of the system to respond to sudden loads and abnormal risks in advance.

[0141] For example, for the conference room scene in the west area of the office building, the time window output by the digital twin simulation is 120 seconds, the adjustment action set includes 3 fan speeds and 2 temperature settings, the energy consumption benchmark is , the comfort score benchmark is , and the time lag benchmark is . The weights are set as , , . For each action a at each time t, the above formula is used to calculate , as the target performance function, which comprehensively evaluates the weighted performance of energy efficiency and comfort at time t and threshold , the sliding window length is , and the step size is . Within the entire simulation time window, the system automatically selects , as the adjustment window with the optimal energy efficiency and comfort, and the corresponding activation threshold is , which triggers the high-speed fan 15s in advance. The final output parameter set for the enhanced learning controller call is {room ID, window [40, 65], instruction a^*=high wind, priority=1}. In actual deployment, the conference room can respond to sudden load changes in demand in advance, reduce energy consumption by 4%, and reduce comfort by less than 1% of the limit, shorten the adjustment response lag to within 8 seconds, and significantly improve the policy linkage and risk prevention capabilities of the system.

[0142] The step S7: input the optimal adjustment timing window and activation threshold into the policy network in the enhanced learning controller, and dynamically optimize the policy network according to the three-element adaptive adjustment mechanism of "expected energy consumption benefit-policy execution lag-comfort tolerance", to output the pre-execution control instruction of the fan coil. Specifically, it includes: 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 input as enhanced learning controller inputs to ensure that the policy network can integrate the dynamic deduction results of the simulation unit on the fan coil load trend and device abnormal risk to adjust the regulation and control strategy adaptation parameters at the current time.

[0143] S7.2: Joint processing of the input optimal adjustment timing window, activation threshold, and load mutation risk value and abnormal tendency characteristics generated by the multi-objective neural network timing prediction model, through an adaptive weight distribution algorithm in the policy network structure, the weight parameters of the three performance indicators of 'expected energy consumption benefit','strategy execution lag', and 'comfort tolerance' in the decision-making process are distributed to realize dynamic trade-off in multi-objective optimization.

[0144] For the input optimal adjustment timing window, activation threshold, and load mutation risk value and abnormal tendency characteristics generated by the multi-objective neural network timing prediction model, a multi-source heterogeneous data fusion method is used (parameters: timing window boundary, threshold parameter, load risk vector, abnormal probability characteristics) to realize the structured input of different types of control parameters in a unified decision space.

[0145] An adaptive weight distribution algorithm in the policy network (parameters: expected energy consumption benefit target, strategy execution lag tolerance, comfort tolerance threshold) is used to establish a three-dimensional performance indicator mapping relationship for the fused parameter set, and 'expected energy consumption benefit','strategy execution lag', and 'comfort tolerance' are embedded as independent performance dimensions in the policy network weight layer.

[0146] Further, through a multi-objective optimization constraint mechanism, a joint loss function of the policy network performance target is defined, and the three-dimensional performance indicators are dynamically weighted using the following weight distribution formula: wherein, is the expected energy consumption benefit weight, is the comfort tolerance weight, is the strategy execution lag weight, and the sum of the three is 1, is the weighted comprehensive performance score.

[0147] is the energy consumption change amount brought by the action, is the comfort improvement amount, is the response lag introduced by the control action, , , is the normalized reference value.

[0148] Adopting a dynamic weight self-adaptive allocation method (parameters: historical strategy execution effect backtracking, mutation risk detection probability, regulation lag trend), based on a periodic rolling window, the expected value, variance and response sensitivity coefficient of each performance indicator are calculated, and the system state and predicted output are dynamically adjusted according to the real-time working condition 、 、 .

[0149] Further, by adopting weight normalization mapping constraints (parameters: normalization reference, extreme value clipping), the weight component of each performance indicator is projected in the interval, and the weight extreme value transition caused by sudden working conditions is forced to be constrained, avoiding the nonlinear instability of the strategy network.

[0150] Through the above chain derivation, the joint dynamic weight allocation for the optimal regulation time window and the activation threshold and related prediction indicators is realized, and the multi-objective optimization target is continuously embedded into the strategy network structure in the form of adjustable parameters, effectively supporting the downstream strategy network dynamic optimization and action selection.

[0151] For example, taking the conference room working condition in office building A as an example, the input optimal regulation time window is [30s, 60s], the activation threshold is high wind speed, the load mutation risk value is 0.85, and the abnormal tendency feature is 0.15. Based on the regulation data backtracking of the last 5 days, the historical cumulative energy consumption reduction rate is 6.5% on average, the comfort score fluctuation is controlled within ±2%, and the average response time lag is 12 seconds. Set the initial weight (focus on energy saving), (modest comfort), (moderate lag risk), according to real-time monitoring, if the predicted load mutation probability is increased to 0.92, the system automatically rolls up to 0.25, and reduces to sacrifice a certain comfort level to shorten the control time lag. After comprehensive allocation of weights, the strategy network dynamically outputs the optimal action distribution under this time window, providing the optimal parameter basis for subsequent value function estimation and action selector. The actual deployment effect verification, energy consumption response curve and comfort score are improved synchronously, and the strategy network shows the ability of immediate adaptation and optimization decision to sudden load and abnormal state.

[0152] S7.3: Using the above adaptive weight parameters, the value iteration calculation is performed on the value function estimation module in the strategy network, so that the regulation window (cause) obtained by simulation deduction is fully considered in the strategy evaluation process. The influence (result) on the final control activation time point and comfort, and the dynamically optimized regulation strategy vector is obtained.

[0153] S7.4: input the dynamically optimized control strategy vector into the action selector in the enhanced learning controller, complete the action probability distribution adjustment based on the soft maximum strategy or the ε-greedy strategy, and output the fan coil pre-execution control instruction that takes into account energy efficiency, time delay and comfort tolerance.

[0154] S7.5: post-consistency check on the output fan coil pre-execution control instruction, real-time correction of the parameter update rate and action threshold in the strategy network based on the current cycle prediction error and strategy execution delay feedback, to ensure that the next cycle strategy optimization can continuously evolve according to the historical causal chain, realizing online adaptive and robust output of the central air conditioning fan coil enhanced learning controller.

[0155] The step S8: according to the pre-execution control instruction, real-time operation command is issued to the central air conditioning fan coil end, and the heating, refrigeration and air speed adjustment instruction is activated in advance to make the system dynamically respond to load mutation or equipment abnormality in prediction. Specifically, it includes: S8.1: based on the pre-execution control instruction obtained in the previous step, analyze the specific parameters in the control instruction, including target air speed, expected temperature set value and execution priority label, to obtain a complete set of fan coil end operation command, to clearly assign data values to each control entry of the device.

[0156] S8.2: execute state discrimination on the fan coil end, obtain the fan running state, valve opening and closing state and in-place feedback of the current end device, check the device linkage condition through the device state synchronization mechanism, and ensure that the time window of the operation command issuance matches the actual controllable state of the device, to optimize the execution timing of the downstream execution action.

[0157] S8.3: use field-level communication protocols (such as Modbus, BACnet, etc.) to real-time issue the generated fan coil end operation command to the corresponding fan coil controller of each room in batches and according to priority through an encrypted command channel, to ensure that the control instruction is fully covered and has anti-interference ability.

[0158] S8.4: implement instruction mapping on the fan coil controller side, map the received operation command to hardware action instruction through the device driver layer, realize hardware-level trigger control of heating function, refrigeration function and air speed gear, and complete the physical action closed loop.

[0159] Take the fan coil end operation command set as the input data, including target air speed instruction, expected temperature set value and control priority parameter, and require to complete the mapping of the instruction to the physical action on the fan coil controller side to ensure short-time response.

[0160] An instruction analysis protocol layer module (parameters: operation command format, function enumeration table, instruction decomposition rule) is adopted to realize classification and analysis of the received fan coil end operation command set, to distinguish heating, refrigeration and wind speed adjustment and other control functions, and to accurately extract key control parameters and action priority identifiers.

[0161] Further, through a device driver mapping adapter (parameters: driver interface dictionary, hardware adaptation layer, control mapping table), the parsed logical commands are converted into corresponding bottom layer hardware action codes according to the driver layer mapping rules, to form level signals, PWM adjustment signals and field bus data formats compatible with terminal execution devices, to ensure smooth linkage of operation commands and fan coil controller hardware control units.

[0162] Further, a multi-channel hardware scheduling mechanism (parameters: heating relay control port, refrigeration valve execution port, wind speed PWM signal channel, priority queue) is adopted to implement parallel or time-sharing scheduling of each function control channel according to the priority order set by the operation command, so that heating, refrigeration and wind speed adjustment instructions can be triggered independently or in combination, while preventing instruction conflicts and hardware level deadlocks.

[0163] Further, through a state monitoring and action closed loop detection module (parameters: action feedback loop, in-place detection signal, timeout protection threshold), action response signals of each physical execution component of the fan coil are collected in real time to determine whether each function operation has been accurately executed, and an alarm is issued in abnormal cases to ensure hardware level safety and execution consistency.

[0164] Through the above-mentioned driver mapping and closed loop control implementation modules, the parsed central air conditioning fan coil end operation commands are efficiently converted into hardware action instructions, to complete precise physical trigger control of heating, refrigeration and wind speed gears, to realize physical action closed loop of the central air conditioning fan coil, and to support real-time adjustment requirements of the upper layer intelligent regulation and control.

[0165] For example, for the fan coil unit in the three-story east conference room in Zone A of Building A, the controller continuously monitors the risk of load mutation and, in the kth cycle, issues a high-priority pre-execution operation command through the prediction model, requiring the wind speed to be increased to the maximum (wind speed setting value = 3), the temperature to be set to 21°C, and the refrigeration function to be activated. The controller-side instruction analysis protocol layer splits the operation command structure into wind speed level instructions, temperature control target instructions, and refrigeration function bits according to the set command template, and maps them to the PWM control channel (duty cycle 100%), the temperature closed-loop regulation port (reference voltage 0.7V), and the refrigeration valve output port (high level 9V), respectively. The device driver layer triggers the wind speed relay, the temperature regulation module, and the refrigeration valve motor in sequence according to the hardware mapping table, listens to the in-place signals, detects that all operation actions are completed within 2.5 seconds, and feeds back the status signal as "execution success". The system records the action response curve, measures the room temperature change gradient after the control action as 0.22°C / min, and the wind speed increase response time delay as 1.7 seconds, which meets the set indicators of the intelligent control system for response speed and accuracy. Multiple measurements show that the fan coil unit controller side using the above chain instruction mapping and hardware layer closed loop can stably realize the pre-trigger control action in the load mutation scenario, and the actual physical effect is highly consistent with the prediction target.

[0166] S8.5: Real-time collection of feedback data of the fan coil unit end after executing the operation command, including action response time delay, temperature control change curve, and actual adjustment amount of wind speed, reporting the running state changes to the control master station through the control instruction execution feedback mechanism to realize the feedback data link closed loop.

[0167] The step S9: periodically monitor the actual feedback data after control, compare the actual energy consumption change, comfort level, and abnormal occurrence dynamics with the aforementioned prediction results, optimize the strategy network and migrate the model parameters through self-evolution learning, and realize the predictive intelligent control closed loop of the central air conditioning fan coil unit. Specifically, it includes: S9.1: Collect the energy consumption feedback data after periodic control, use the energy consumption sensor network to obtain the actual energy consumption parameter value from each fan coil unit end in real time as the system feedback input to support subsequent control effect analysis.

[0168] S9.2: Collect the comfort feedback data after periodic control, use the indoor environment sensor array to obtain the room temperature and humidity response changes, and use them as feedback inputs to represent the actual influence of the control action on user comfort.

[0169] S9.3: Collect the equipment abnormality dynamics after periodic control, capture abnormal features such as fan current, valve opening, and equipment surface temperature based on the fault detection algorithm, realize real-time monitoring of the equipment state, and form equipment abnormality occurrence dynamic feedback data.

[0170] S9.4: Compare the actual energy consumption feedback parameters, comfort feedback parameters, and equipment abnormality feedback parameters with the multi-objective neural network prediction model output results, and use residual analysis and trend difference quantification methods to obtain the accuracy index of the model prediction and the deviation measurement of the current control strategy.

[0171] S9.5: According to the actual feedback data and the prediction residual, trigger the self-evolution learning mechanism, and use the adaptive gradient optimization algorithm to fine-tune the reinforcement learning strategy network weight parameters to further adapt to the actual working condition changes and optimize the control decision accuracy.

[0172] S9.6: Combine the working condition residual characteristics and the historical learning curve to automatically adjust the parameter prior weight and the migration strategy in the cross-domain knowledge migration module to improve the self-adaptive ability of the migration model to new scenes, dynamic loads, and equipment aging environments.

[0173] S9.7: Periodically synchronize the optimized reinforcement learning strategy network parameters and the knowledge migration module parameters to the central air conditioning fan coil predictive intelligent control system master unit to complete the closed-loop self-evolution control logic and ensure the self-adaptive evolution and continuous optimization of the system in various working condition environments.

[0174] For those skilled in the art, various corresponding changes and modifications can be made to the above-described technical solutions and concepts, and all such changes and modifications should be within the scope of protection of the claims of the present application.

[0175] Unless otherwise defined, technical terms or scientific terms used herein should be understood as having the common meaning to those skilled in the art. The terms "first", "second", "third" and similar terms used in the patent application specification and claims do not represent any order, number or importance, but are only used to distinguish different components. Similarly, the terms "one" or "a" and similar terms do not represent a quantity limitation, but represent the existence of at least one. The terms "include" or "contain" and similar terms mean that the elements or objects appearing before "include" or "contain" cover the elements or objects listed after "include" or "contain" and their equivalents, and do not exclude other elements or objects. The multiple involved in the embodiments of the present application refers to two or more. A and / or B means that there are three cases: A; B; and A and B.

[0176] The above description is only an exemplary embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A central air conditioning fan coil intelligent regulation method, specifically comprising: S1: collecting environmental parameter data of multiple positions at the end of the central air conditioning fan coil; S2: performing time series denoising, normalization and outlier rejection processing on the environmental parameter data to obtain a preprocessed data set; S3: based on the preprocessed data set, extracting load change characteristics, comfort indicators and energy consumption trend characteristics for different room types, regional attributes and equipment aging states, and generating multi-dimensional working condition labels; S4: inputting the multi-dimensional working condition labels into a cross-domain knowledge transfer model, using the energy consumption and comfort adjustment parameters learned in the known scene to perform knowledge transfer and parameter pre-initialization to obtain the parameter prior of the current target environment; S5: predicting the load trend and equipment abnormality probability of the room in a future preset time window based on a neural network time series prediction model combined with internal sensor data and external dynamic data, outputting a load mutation risk value and an abnormality tendency feature; S6: constructing a real-time digital twin simulation unit, virtually calculating the consequences of executing various adjustment strategies according to the parameter prior and mutation risk value, predicting the influence of fan coil adjustment actions on temperature control response and energy consumption, and determining the optimal adjustment time window and activation threshold; S7: inputting the optimal adjustment time window and activation threshold to dynamically optimize the policy network in the reinforcement learning controller, and outputting the pre-execution control instructions of the fan coil.

2. The intelligent control method for fan-coil units of a central air conditioner according to claim 1, characterized in that, The environmental parameter data includes room temperature and humidity, fan current, valve opening, equipment surface temperature, room traffic statistics, external meteorological data and building direction information.

3. The intelligent control method of a fan coil unit of a central air conditioner according to claim 1, characterized in that, After step S7, step S8 is further included: according to the pre-execution control instructions, real-time operation commands are issued to the end of the central air conditioning fan coil to activate the heating, refrigeration and air speed adjustment instructions in advance, so that the system dynamically responds to the load mutation or equipment abnormality in the prediction.

4. The intelligent control method of a fan coil unit of a central air conditioner according to claim 3, characterized in that, After step S8, step S9 is further included: periodically monitor the actual feedback data after regulation and control, compare the actual energy consumption change, comfort level and abnormality occurrence dynamics with the aforementioned prediction results, and optimize the policy network and transfer model parameters through self-evolution learning.

5. The intelligent control method of a fan coil unit of a central air conditioner according to claim 2, wherein, The external meteorological data set integrates external meteorological parameters such as weather temperature, humidity, wind speed and air pressure.

6. The intelligent control method for central air-conditioning fan coil units according to claim 1, characterized in that: Step S4 specifically includes: In the multi-source information fusion process, the multi-dimensional working condition labels generated are normalized using a label standardization algorithm based on the regional attributes, room types and equipment aging states corresponding to the labels, to obtain label feature vectors suitable for cross-domain transfer; Using the standardized label feature vectors as input, and using the energy consumption characteristics parameters and comfort adjustment parameters accumulated in the historical known scene as the source domain, a set of prior adjustment parameters with high correlation to the current target label is extracted; The prior adjustment parameters are redefined and adaptively weighted to eliminate the potential distribution deviation of the source domain and the target environment in energy consumption regulation and comfort indicators; For the obtained parameter prior, the label feature vectors and room dynamic working condition features are used to cluster, archive and select key parameters. The parameter set obtained through the final screening is input into the neural network time series prediction model and the digital twin simulation unit to initialize the model parameters and the simulation scene settings, thereby achieving dynamic adaptive initialization and cross-scene control capability improvement at the system level.

7. The intelligent control method of a fan coil unit of a central air conditioner according to claim 1, characterized in that: The S1 comprises using a distributed room temperature and humidity sensor network to obtain temperature and relative humidity data of multiple sampling points in the room to form a structured temperature and humidity parameter matrix.

8. The intelligent control method of a fan coil unit of a central air conditioner according to claim 1, characterized in that: The time series window length of the digital twin simulation output is 120 seconds.

9. The intelligent control method of a fan coil unit of a central air conditioner according to claim 1, characterized in that: The S2 comprises using an adaptive sliding filter algorithm to suppress time series noise for the collected multi-channel environmental parameter data, and using Z-score standardization or Min-Max normalization to perform channel-by-channel normalization processing.

Images