Intelligent energy-saving optimization method and device for variable-frequency centrifugal fan

Abstract

The invention relates to the field of energy-saving optimization of variable-frequency fans, in particular to an intelligent energy-saving optimization method and device for a variable-frequency centrifugal fan. The method comprises the following steps: acquiring environmental monitoring parameters in a region, performing dynamic operation scene analysis, performing environmental parameter distribution change mining, and constructing a dynamic scene operation map; obtaining a manual operation log of a historical centrifugal fan, carrying out user use preference mining and equipment response matching, and constructing a personalized fan use behavior portrait; based on the dynamic scene operation map and the personalized fan use behavior portrait, fan use rule mining in the building is carried out, fan operation demand prediction is carried out, and a real-time fan load demand map is generated; real-time operation parameters of the centrifugal fan are collected, principal component dimension reduction analysis and self-adaptive feature coding are carried out, and a fan real-time operation condition matrix is constructed. Fine energy efficiency control is achieved, the energy consumption needed by unit air volume is remarkably reduced, and the energy-saving efficiency of the draught fan and the user experience are improved.

Description

Technical Field

[0001] The present invention relates to the field of energy-saving optimization of variable-frequency fans, and in particular to an intelligent energy-saving optimization method and device for variable-frequency centrifugal fans. Background Art

[0002] With the continuous advancement of urbanization and increasing demands for living comfort, the demand for ventilation systems in various public buildings, commercial complexes, and residential communities continues to grow. As a core component of building ventilation systems, variable-frequency centrifugal fans, due to their high efficiency, adjustable air volume, and low noise levels, have been widely used in various building areas for air conditioning, ventilation, smoke exhaust, and ventilation. However, in actual operation, variable-frequency centrifugal fans are often affected by factors such as system load fluctuations, external environmental changes, and inappropriate control parameters. This can cause them to frequently operate in suboptimal conditions, resulting in energy waste and reduced equipment efficiency. Furthermore, traditional fan control systems generally rely on preset control logic and fixed parameters, lacking dynamic response capabilities and intelligent adjustment mechanisms, making them unable to adapt to the complex and changing ventilation needs within a building area. Furthermore, current energy-saving control methods often rely on manually set operating strategies, resulting in adjustment lags, making it difficult to achieve real-time system optimization and adaptive energy efficiency control, thus hindering overall fan performance. Therefore, it is urgent to propose a variable frequency centrifugal fan energy-saving control method for building areas that integrates intelligent analysis, real-time regulation and energy-saving optimization, so as to achieve high efficiency, economy and intelligence in the operation of building ventilation systems and promote the overall improvement of building energy management level. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention proposes an intelligent energy-saving optimization method and device for a variable frequency centrifugal fan to solve at least one of the above technical problems.

[0004] To achieve the above object, the present invention provides an intelligent energy-saving optimization method for a variable frequency centrifugal fan, comprising the following steps: Step S1: Collect environmental monitoring parameters in the area, perform dynamic operation scenario analysis, and mine the distribution changes of environmental parameters to build a dynamic scenario operation map; Step S2: Obtain historical centrifugal fan human operation logs, conduct user usage preference mining and device response matching, and build a personalized fan usage behavior profile; Step S3: Based on the dynamic scenario operation map and personalized fan usage behavior portrait, the fan usage pattern in the building is mined, and the fan operation demand is predicted to generate a real-time fan load demand map; Step S4: collecting the real-time operating parameters of the centrifugal fan, performing principal component dimensionality reduction analysis and adaptive feature coding, and constructing a real-time operating condition matrix of the fan; Step S5: performing a multi-objective energy-saving constraint analysis on the real-time operating condition matrix of the fan based on the real-time fan load demand diagram, and performing a minimum energy-saving calculation to obtain the minimum energy-saving control parameters; Step S6: Execute dynamic control of the fan based on the minimum energy-saving control parameters, perform dynamic closed-loop iterative correction, and build an intelligent energy-saving iterative optimization engine.

[0005] In this specification, an intelligent energy-saving optimization device for a variable frequency centrifugal fan is provided, which is used to execute the intelligent energy-saving optimization method for a variable frequency centrifugal fan as described above, including: The scenario analysis module is used to collect environmental monitoring parameters in the area, perform dynamic operation scenario analysis, and mine the distribution changes of environmental parameters to build a dynamic scenario operation map; The usage behavior module is used to obtain historical centrifugal fan human operation logs, conduct user usage preference mining and device response matching, and build a personalized fan usage behavior profile; The demand forecasting module is used to mine fan usage patterns within buildings based on dynamic scenario operation graphs and personalized fan usage behavior portraits, and to forecast fan operation demand and generate real-time fan load demand graphs; The fan operating condition analysis module is used to collect the real-time operating parameters of the centrifugal fan, perform principal component dimensionality reduction analysis and adaptive feature coding, and construct the fan real-time operating condition matrix; Energy-saving constraint analysis module, which is used to perform multi-objective energy-saving constraint analysis on the real-time operating condition matrix of the fan based on the real-time fan load demand diagram, and perform minimum energy-saving calculation to obtain the minimum energy-saving control parameters; The energy-saving iterative optimization module is used to perform dynamic control of the fan based on the minimum energy-saving control parameters, and perform dynamic closed-loop iterative correction to build an intelligent energy-saving iterative optimization engine.

[0006] The beneficial effects of the present invention are as follows: by monitoring regional environmental parameters (such as temperature and humidity, carbon dioxide concentration, By combining multi-dimensional data collection and time series modeling (e.g., data sources like air flow and occupancy density), dynamic time warping, and spatial thermal distribution analysis, this system achieves a multi-scale deconstruction of operating scenarios. By constructing a dynamic operating map, the spatiotemporal evolution of building internal environmental loads can be characterized, providing high-dimensional environmental driver input for fan operation, thereby improving the scenario-based adaptability of subsequent load prediction and control strategies. Based on historical centrifugal fan control command logs, a highly accurate personalized behavior profile model is constructed by mining temporal behavior patterns, user preference distribution, and control response relationships. This model, combining cluster analysis with Bayesian inference methods, accurately captures users' implicit control preferences for air volume, start frequency, and response delay, significantly improving the subsequent control algorithm's ability to identify human intervention intentions and achieving coordinated optimization of "user intent-driven" and "system energy-saving goals." After integrating the dynamic scenario operating map with personalized user behavior profiles, the system then models and predicts fan operating patterns using time series modeling (such as LSTM or Transformer), generating a granularly adjustable fan load demand map. This process can not only capture potential load fluctuations in advance, but also provide predictive priors for subsequent control optimization, introduce a feedforward mechanism in the control strategy design, thereby reducing response lag and improving control accuracy and the foresight of energy consumption scheduling.

[0007] After collecting real-time operating parameters of centrifugal fans (such as current, voltage, air pressure, air volume, speed, and vibration), various dimensionality reduction techniques, such as principal component analysis (PCA) and independent component analysis (ICA), are used to compress redundant features. Adaptive encoding (such as autoencoders or feature embedding) is then combined to extract key operating conditions. The resulting operating condition matrix exhibits high discriminability and low-dimensional representation, reducing computational complexity while improving sensitivity to operational state changes (such as underload, overload, and fault trends). This serves as the core input for subsequent multi-objective optimization. Based on the fan load forecast map and the real-time operating condition matrix, a multi-objective energy-saving optimization model is constructed. This model incorporates energy consumption functions, comfort constraints, and response rate for joint optimization. Strategies such as nonlinear programming (NLP) and particle swarm optimization (PSO) are employed to determine the optimal control parameters for the current scenario. This parameter set corresponds to the theoretical minimum energy consumption point under the fan's current operating conditions, significantly reducing energy consumption per unit air volume, enabling refined energy efficiency control, and providing practical optimization solutions for dynamic control strategies. By implementing minimum energy-saving control parameters and combining real-time wind turbine feedback data with environmental changes, a dynamic control mechanism based on model predictive control (MPC) or reinforcement learning (RL) is constructed to achieve closed-loop iterative correction of the control strategy. The system adaptively adjusts based on energy-saving results, system response deviations, and environmental disturbances, building an intelligent energy-saving iterative optimization engine that enables continuous learning and optimization of the wind turbine system, ultimately achieving the optimal solution for energy saving, stability, and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1A schematic flow chart of the steps of an intelligent energy-saving optimization method for a variable frequency centrifugal fan according to the present invention; Figure 2 Detailed implementation flow chart of step S1; Figure 3 Detailed implementation flow chart of step S2; Figure 4 Schematic diagram of the detailed implementation steps of step S3. DETAILED DESCRIPTION

[0009] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0010] This application provides an intelligent energy-saving optimization method and device for a variable-frequency centrifugal fan. The execution entities of the intelligent energy-saving optimization method and device for a variable-frequency centrifugal fan include, but are not limited to, the following: mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc. equipped with the system, which can be regarded as general computing nodes of this application. The data processing platform includes, but is not limited to, at least one of an audio and image management system, an information management system, and a cloud data management system.

[0011] See also Figures 1 to 4 The present invention provides an intelligent energy-saving optimization method for a variable frequency centrifugal fan, comprising the following steps: Step S1: Collect environmental monitoring parameters in the area, perform dynamic operation scenario analysis, and mine the distribution changes of environmental parameters to build a dynamic scenario operation map; Step S2: Obtain historical centrifugal fan human operation logs, conduct user usage preference mining and device response matching, and build a personalized fan usage behavior profile; Step S3: Based on the dynamic scenario operation map and personalized fan usage behavior portrait, the fan usage pattern in the building is mined, and the fan operation demand is predicted to generate a real-time fan load demand map; Step S4: collecting the real-time operating parameters of the centrifugal fan, performing principal component dimensionality reduction analysis and adaptive feature coding, and constructing a real-time operating condition matrix of the fan; Step S5: performing a multi-objective energy-saving constraint analysis on the real-time operating condition matrix of the fan based on the real-time fan load demand diagram, and performing a minimum energy-saving calculation to obtain the minimum energy-saving control parameters; Step S6: Execute dynamic control of the fan based on the minimum energy-saving control parameters, perform dynamic closed-loop iterative correction, and build an intelligent energy-saving iterative optimization engine.

[0012] In the embodiment of the present invention, see Figure 1, is a schematic flow chart of the steps of an intelligent energy-saving optimization method for a variable frequency centrifugal fan according to the present invention. In this example, the steps of the intelligent energy-saving optimization method for a variable frequency centrifugal fan include: Step S1: Collect environmental monitoring parameters in the area, perform dynamic operation scenario analysis, and mine the distribution changes of environmental parameters to build a dynamic scenario operation map; In this example, high-precision multimodal sensors (such as temperature and humidity sensors, carbon dioxide concentration sensors, PM2.5 particulate matter sensors, and light intensity sensors) are deployed at multiple key locations within the building, including meeting rooms, open office areas, corridors, and equipment rooms. Each sensor reports data at a frequency of 5 to 30 seconds, continuously recording environmental parameters. To ensure temporal consistency and spatial coverage, edge computing devices (such as the Jetson Nano) are used to aggregate, denoise, and timestamp the raw data, resulting in a continuous, stable, and comparable monitoring data stream.

[0013] The collected multi-dimensional environmental parameter data is mapped to the three-dimensional model of the building space, and the semantic annotation of each functional area is carried out in combination with the building BIM (Building Information Modeling) model. Cluster analysis (such as DBSCAN) and principal component analysis (PCA) methods are used to aggregate the environmental characteristics of different spatial areas in the same time period, thereby achieving a preliminary division of the regional operating status (such as high temperature and high humidity, low temperature and low humidity, dense population, high concentration). In an experimental sample, this method identified 15 typical scenarios in an office building and tracked their dynamic change paths using labels. Mining environmental parameter distribution changes further focuses on identifying changing trends and sudden changes in environmental factors over time. Parameter sequences are statistically analyzed using a sliding time window method (e.g., a 60-minute period). Time series anomaly detection techniques (such as the Prophet model and the Local Outlier Factor (LOF) algorithm) are used to identify short-term abnormal fluctuations in environmental parameters. For example, if indoor carbon dioxide concentration rapidly rises from 550 ppm to 1100 ppm within 20 minutes, the system identifies this as a "localized intensive use anomaly." Such sudden changes directly impact fan speed regulation and energy consumption strategies. The results of this multidimensional analysis are structured as a graph using a graph database (such as Neo4j). Nodes represent functional areas and environmental states, edges represent transitions or associations between states, and attributes annotate various monitoring data and statistical values. This constructs a dynamic scenario operation graph. This graph not only supports real-time monitoring and tracking of environmental conditions but also provides intuitive and traceable contextual evidence for subsequent optimization of fan response strategies.

[0014] Step S2: Obtain historical centrifugal fan human operation logs, conduct user usage preference mining and device response matching, and build a personalized fan usage behavior profile; In this embodiment, user operation log data from the past 6 to 12 months is retrieved from the wind turbine control system (such as a PLC or BAS system). This data mainly includes operation information such as wind turbine start and stop time, operating mode switching (such as manual / automatic), frequency adjustment instructions, and manual energy-saving mode on / off. Each record must be associated with fields such as a timestamp, operating user ID (which can be obtained through the login password or permission management module), operation location, and corresponding device number. After obtaining the operation log, data preprocessing is first performed, mainly including format unification, missing data filling (nearest neighbor interpolation can be used), duplicate record cleaning, and time alignment. Next, an anomaly detection algorithm (such as Isolation Forest) is used to preliminarily screen out abnormal operating behaviors in the log, such as frequent start and stop operations in unmanned areas at night, to ensure that the preference data subsequently mined is representative and reliable.

[0015] The user operation behavior identification phase begins. Sequence analysis methods (such as n-gram behavior models and Markov chains) are used to analyze each user's operation sequences over different time periods. The impact of these operations is assessed by combining them with actual device response data (such as real-time changes in fan frequency, power, and temperature rise). For example, in the experimental sample, some users habitually set the fan to 80% load frequency before 9:00 a.m., then manually adjust it to 60% in the afternoon to account for changes in occupancy and rising temperatures. By statistically analyzing these frequently occurring behavior patterns, representative sets of behavior patterns can be further extracted. Cluster analysis techniques (such as K-Means, DBSCAN, or GMM) are used to categorize user preference profiles, grouping users with similar operating styles into several categories, such as "energy-saving priority," "comfort priority," and "frequent intervention." Based on log data from 50 actual operators, the system identified six distinct groups of user preferences, showing significant differences in start / stop frequency, frequency setting range, and manual intervention time. The user preference model is linked to the device response model to construct a personalized fan usage behavior profile. This profile includes not only static characteristics (such as common operating hours and preferred frequency ranges) but also dynamic response characteristics (such as device stabilization time after frequency changes and energy consumption response). All profile information is stored in a structured form in the behavioral profile database and can be used for subsequent adaptive wind turbine control and personalized energy-saving strategy matching.

[0016] Step S3: Based on the dynamic scenario operation map and personalized fan usage behavior portrait, the fan usage pattern in the building is mined, and the fan operation demand is predicted to generate a real-time fan load demand map; In this embodiment, the dynamic scene operation map constructed in the previous step is used, which integrates the temperature and humidity distribution of multiple time periods and across regions inside the building, The system captures multimodal environmental information, including concentration changes, occupant density, and behavioral trajectories. Based on the spatiotemporal environmental evolution characteristics encoded in this atlas, the system uses a sliding time window mechanism (e.g., a 15-minute window step) to analyze environmental pressure trends in different areas at different times. For example, a surge in occupancy around noon on an office floor can lead to a localized increase in CO2 concentration, which in turn triggers an increase in fan load demand. Simultaneously, typical operational behavior sequences for the target user group are extracted from the behavioral profile database and mapped and integrated with these operational patterns. This process employs multimodal association rule mining algorithms (e.g., the extended Apriori model) to identify linkage patterns between user operational behavior and environmental changes, such as the behavioral pattern of "increased CO2 concentration + increased user operation frequency → increased fan frequency setting." The experiment used log data from five working days within a complex building and found that over 80% of fan frequency increases during peak hours were directly associated with excessive occupancy and elevated CO2 concentrations. Having identified the dual evolutionary patterns of the scene environment and user behavior, the system then predicts future fan operating demand. The prediction method uses a fusion of time series modeling and deep learning methods. Typical models include LSTM (Long Short-Term Memory Network) combined with Attention mechanism. Input features include temperature and humidity in the current and past n hours, , occupancy density sequences, historical operational behavior sequences, energy-saving control records, etc., and outputs the fan operating frequency and estimated load every 10 minutes for the next hour. To improve the model's real-time performance and generalization capabilities, an autoregressive sliding window validation mechanism was introduced. This means that after the model's prediction, new measured load data is fed back into the training set for lightweight iterative model updates. In the experiment, the fan prediction system deployed in a multi-functional office building achieved a prediction accuracy (measured by RMSE) exceeding 95%, with a prediction error within ±5Hz. The system visualizes the predicted future load demand for each area and fan in a graphical form, constructing a real-time fan load demand map. This map is based on the building floor plan and overlaid with the dynamic load thermal distribution. Colors indicate load intensity, and the timeline allows users to scroll through prediction results within different time windows, providing intuitive decision support for subsequent adaptive fan scheduling and energy optimization.

[0017] Step S4: collecting the real-time operating parameters of the centrifugal fan, performing principal component dimensionality reduction analysis and adaptive feature coding, and constructing a real-time operating condition matrix of the fan; In this embodiment, real-time operating parameters of variable-frequency centrifugal fans are collected at high frequency and across all dimensions to construct a comprehensive data structure covering electrical, mechanical, thermal, and vibration aspects. Key collected parameters include: three-phase current, voltage, operating frequency, motor speed, impeller shaft vibration signals (lateral and longitudinal), motor housing temperature, outlet air pressure, and air volume trends. The system deploys high-precision energy meters and temperature sensors, and multi-axis accelerometers are installed at key locations on the fan housing to capture high-frequency mechanical vibration signals. The sampling frequency is set to 1000Hz for electrical parameters (current, voltage, and frequency), 1Hz for mechanical and thermal parameters (temperature and speed), and 10kHz for vibration signals for subsequent spectral feature analysis. The collected raw data undergoes preprocessing in an edge computing module, including normalization, noise filtering (wavelet denoising), time alignment, and outlier identification, to ensure the constructed data is analytically valuable and timely. For example, for three fans in an industrial plant, the amount of data collected every 24 hours exceeds 15GB, forming a critical foundation for building a data-driven fan operation model. Principal component dimensionality reduction analysis (PCA) is required to reduce redundancy, compress feature dimensions, and retain the maximum amount of information. PCA extracts the top principal components with the largest variance in the data by performing eigenvalue decomposition on the covariance matrix between parameters, thereby achieving information compression and dimensionality reduction. In actual operation, a total of 18 first-order and second-order features, including vibration RMS value, current mean, speed fluctuation amplitude, temperature rise rate, and number of frequency jumps, were selected as input variables. After PCA processing, the top 6 principal components were extracted. These principal components cumulatively explained 91.3% of the variance of the original data, effectively retaining the core change information in the operation of the wind turbine. In addition, the system sets up a dynamic update mechanism to regularly recalculate the principal component load matrix to adapt to changes in long-term operating conditions. Through this processing, redundant or highly correlated features are effectively eliminated, which significantly improves the computational efficiency and stability of subsequent models.

[0018] The reduced-dimensional data then enters the adaptive feature encoding (AFE) stage, which uses nonlinear encoding techniques to enhance feature representation and form a unified input structure. Specifically, an autoencoder architecture is introduced to compress and reconstruct the principal component features, enabling the model to automatically learn the nonlinear mapping relationship between features. The encoder employs a three-layer fully connected network with a Reluctant Unit (ReLU) activation function. The encoding dimensions are ultimately compressed to 3-4 dimensions, with reconstruction loss kept to an average significance error (MSE) of less than 0.002. This encoding not only preserves the global trends of operating characteristics but also improves the ability to identify minor anomalies and precursors to sudden changes. The system also incorporates a feedback-based adaptive mechanism. When a sustained deviation in the wind turbine's operating status is detected (e.g., an increase of more than 10% in the average vibration amplitude), the encoder weights are automatically fine-tuned to ensure robust encoding. This process achieves highly aggregated and sensitive wind turbine status features. A real-time wind turbine operating condition matrix is ​​constructed, with time as the horizontal axis and the encoded principal component features as the vertical axis. This matrix can be viewed as a multidimensional time-evolution graph of wind turbine operation. Each row represents the comprehensive operating state of the wind turbine at a specific moment, and each column represents a feature dimension after dimensionality reduction encoding. This matrix not only provides real-time input for subsequent wind turbine load prediction and energy-saving control strategy generation, but can also be used for a variety of advanced tasks such as wind turbine anomaly identification, state assessment, and operating condition classification.

[0019] Step S5: performing a multi-objective energy-saving constraint analysis on the real-time operating condition matrix of the fan based on the real-time fan load demand diagram, and performing a minimum energy-saving calculation to obtain the minimum energy-saving control parameters; In this embodiment, the constructed real-time fan load demand diagram and the fan real-time operating condition matrix are used as core input data. Among them, the load demand diagram reflects the target operating level that the fan should achieve in different time periods (such as air volume, speed, pressure response, etc.), while the operating condition matrix provides a multi-dimensional feature description of the current operating state of the fan (such as speed, current, frequency, temperature rise, vibration, etc.). The combination of these two data structures constitutes the mapping basis between the current operating state and the target operating demand, which is the prerequisite for carrying out energy-saving optimization analysis. A multi-objective energy-saving constraint analysis method is used to reasonably constrain the operating parameters. The analysis process comprehensively considers the following key objectives: (1) Minimization of energy consumption, that is, the power input per unit time is as low as possible; (2) Response performance guarantee, that is, the fan can output the target air volume and pressure on demand to maintain comfort and system stability; (3) Equipment life protection, that is, avoiding frequent start-stop, high-frequency jump or long-term full-load operation and other behaviors that are more damaging to the equipment. When developing the constraint model, typical operating condition constraints were incorporated, such as maximum allowable current (e.g., no more than 95% of the rated value), minimum speed limit (typically above 30Hz to ensure airflow), and frequency change rate limit (e.g., no more than 10Hz per minute). Environmental adaptability conditions were also established, such as appropriately lowering the upper operating limit in high temperature and high humidity conditions to prevent equipment overheating. In experiments, based on historical operating data from three wind turbines under varying loads during peak hours (9:00–11:00 AM daily) and off-peak hours (nighttime), the established constraint model covered an average of over 95% of actual operating conditions, demonstrating strong adaptability.

[0020] Based on the above constraints, the system determines the optimal operating parameter combination for the wind turbine using a minimum energy-saving calculation model. This calculation method, centered on an optimization algorithm, employs either a multi-objective particle swarm optimization (MOPSO) or differential evolution (DE) algorithm. These algorithms search for optimal solutions in a high-dimensional parameter space and can simultaneously address multiple objective functions. Optimization objective functions typically include energy consumption per unit air volume (kWh / m³), current load factor, and speed stability. The system outputs all acceptable optimal solutions as a multi-objective Pareto front solution set and then selects the final control parameters based on operational priorities (e.g., energy conservation or response). To improve solution efficiency, the system incorporates a dynamic threshold convergence mechanism, enabling optimization calculations to be completed within 5 seconds in real-time scenarios, facilitating application in near-real-time control scenarios. For example, in a medium-load scenario, a wind turbine operated at a frequency of 48 Hz and an average energy consumption of 2.8 kWh before optimization. After optimization, the frequency was reduced to 42 Hz, achieving approximately a 14.2% reduction in energy consumption and a 12% improvement in operational stability while meeting air volume requirements.

[0021] The final output of the minimum energy-saving control parameters include: target operating frequency, optimal speed, recommended current load rate, upper limit of wind pressure and air volume setting, frequency adjustment curve, etc. These parameters will serve as the basis for the control system to be issued, directly acting on the inverter and fan actuator to achieve the actual implementation of energy-saving goals. The system has a monitoring mechanism for the execution effect of the control parameters. If it detects that energy consumption has not decreased as scheduled, the frequency fluctuation is too large, or the equipment response is delayed, it will trigger the parameter adaptive adjustment mechanism to further ensure the stability and efficiency of the control strategy. In the multi-day operation experiment, the system deployed with this parameter optimization module increased the average energy saving rate by 9.6% compared with the traditional constant speed operation mode, and the fan operation response time was shortened by 8.3%, which fully verified the practicality and stability of this method in the fan energy-saving optimization scenario.

[0022] Step S6: Execute dynamic control of the fan based on the minimum energy-saving control parameters, perform dynamic closed-loop iterative correction, and build an intelligent energy-saving iterative optimization engine.

[0023] In this embodiment, the generated minimum energy-saving control parameters (including the optimal frequency setpoint, speed range, current load limit, and frequency adjustment curve) are transmitted in real time to the fan control unit (such as a frequency converter or PLC) via an industrial Ethernet interface or Modbus communication protocol, initiating dynamic energy-saving operation control of the fan. During this process, control instructions are refreshed every one minute and dynamically adjusted based on the real-time operating demand map, ensuring that the fan load is closely coupled with the real-time demand within the building. During control execution, the system continuously collects feedback information on the fan's current status, including real-time frequency, current, voltage, air volume output, vibration status, and energy consumption data, forming a continuous time-series feedback data stream. This feedback information is mapped one-to-one with the control parameters through a signal synchronization mechanism, establishing a complete "control-response" mapping relationship. The system then performs time-series response calculations on this data, such as calculating the response delay after frequency changes through first-order differentials, analyzing the current response smoothness using a moving average window, and calculating the vibration fluctuation range, to derive the response parameter characteristics during actual operation. During the experiment, during the operation test of multiple fans in the building, it was found that when the response delay exceeded 4 seconds, the air volume deviation exceeded 7%, and parameter adaptation adjustment was required; while when the operating frequency fluctuation was less than 2Hz and the response current change rate was controlled within the range of ±3%, the energy saving effect and comfort level reached an ideal state.

[0024] Based on the control-response analysis described above, the system uses a predictive model (such as an LSTM-based energy-saving response estimation model) to generate an expected response target under the current control conditions. This target is then compared with the real-time feedback response parameters to generate a target deviation. For example, if the expected energy consumption at a certain frequency setting is 2.1 kWh, but the actual measured value is 2.5 kWh, the target deviation is +0.4 kWh. Alternatively, if the air volume deviation exceeds ±10%, the system will deem the operating strategy to have deviated. These target differences are then fed into the next round of parameter adjustment mechanisms as error signals. The system then combines the response model with historical optimization records to perform reverse adjustments and fine-tune the control parameters. This process employs reinforcement learning or adaptive PID control methods for feedback correction, ensuring that the system approaches the optimal solution with each round, forming a self-learning closed-loop. To prevent system oscillation caused by over-adjustment, the system sets difference thresholds (such as frequency fluctuation <1.5 Hz and energy consumption deviation <5%) and adjustment boundaries to ensure stable and reliable feedback correction. This mechanism continuously improves the effectiveness of energy-saving strategy execution through dynamic closed-loop iteration, forming a self-evolving and self-adaptive intelligent energy-saving iterative optimization engine. The engine aims to minimize energy-saving errors, combining multiple rounds of feedback with operational data to dynamically adjust the optimal control path. Each optimization iteration takes approximately five minutes, allowing continuous learning and strategy iteration throughout the day. Field tests conducted over 30 days of continuous deployment in a commercial complex showed that this iterative optimization engine improved fan energy efficiency (COP) by 13.5% and reduced average energy consumption by 11.8%, while maintaining indoor temperature and humidity fluctuations within ±1.2°C / ±5%RH. This demonstrates its excellent adaptability and energy-saving performance in complex building environments.

[0025] In this embodiment, refer to Figure 2 , is a flowchart of the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: Using multiple sensors within the building area to collect environmental monitoring parameters within the area, including temperature, humidity, and carbon dioxide concentration; and obtaining video surveillance images within the building area; Calculate the density of people in the area based on the video surveillance image, and perform building operation status analysis to obtain the building operation status; Performing multimodal environmental comprehensive perception based on the environmental monitoring parameters, the density of people in the area, and the operating status of the building to generate multimodal environmental comprehensive perception data; Perform spatiotemporal encoding on the multimodal environmental integrated perception data and conduct architectural behavior pattern analysis to obtain architectural behavior patterns; Dynamic operation scenarios are analyzed based on building behavior patterns, and changes in environmental parameter distribution are mined to construct dynamic scenario operation maps.

[0026] In this embodiment, various types of environmental monitoring sensors are deployed in the building area, mainly including temperature and humidity sensors (such as SHT31 or DHT22) and carbon dioxide concentration sensors (such as MH-Z19B or Telaire T6615). These sensors are reasonably deployed according to the area and ventilation characteristics of the building space, usually one group every 50-100 square meters to ensure the spatial representativeness and sampling accuracy of the data. The acquisition frequency is set to once per minute, and the sampled data is transmitted to the edge computing device via the LoRa or ZigBee protocol to ensure low power consumption and high reliability. During the acquisition process, the sensor not only records the instantaneous value of the environment, but also counts the daily average, peak value and fluctuation range. For example, in a certain office building experimental environment, a total of 32 groups of sensor nodes were set up, covering 4 floors of the building, and a total of 15.8°C-29.6°C temperature range and humidity from 30%-85% were collected. The concentration range is 450ppm–1600ppm. All sensor data will undergo preliminary data verification and noise removal through the edge computing gateway and will be stored in a time series database (such as InfluxDB). This data not only reflects the current air quality status of the area, but also provides an initial environmental baseline for subsequent judgments on the intensity of personnel activities and the intensity of equipment operation. High-definition video surveillance equipment is deployed in key areas of the building (such as conference rooms, corridors, office areas, etc.) with a resolution of at least 1080p and a frame rate of 15–25fps to ensure sufficient video details for subsequent analysis. The camera angle design needs to avoid occlusion, and the layout height is usually between 2.5–3 meters to obtain a bird's-eye view, which helps to accurately detect and track human targets. Based on the collected video images, advanced deep learning target detection models such as YOLOv5 or YOLOv7 are used for personnel identification, and then combined with background modeling and time series analysis technology to eliminate duplicate counting and stationary target errors, and finally calculate the personnel density per unit area (unit: person / ㎡). In the experiment, we divided the space into 0.5 The system uses a sliding window averaging technique (with a 10-second window size) to generate a stable density estimate. For example, in a typical meeting scenario, the peak density reaches 1.8 people / m², while the average density in office areas is 0.3–0.6 people / m². Furthermore, regional segmentation techniques (such as DBSCAN-based crowd gathering detection) are introduced to enable spatially hierarchical density analysis, providing a more granular basis for human activity analysis in subsequent operational status modeling and energy-saving strategies.

[0027] Based on the collected environmental monitoring data and personnel density data, the building operation status model is constructed. Relying on the multivariate fusion analysis method, the Bayesian network and the decision tree are mainly used to realize the classification and identification of the operation status. The operation status includes four categories: "normal operation", "high load operation", "low load operation" and "abnormal operation". First, a model containing environmental parameters (temperature, humidity, Principal Component Analysis (PCA) was used to reduce the multidimensional feature vectors of CO2 concentration and occupancy density to improve model efficiency. In the experiment, the model was trained using 500 hours of field data, with thresholds for abnormal environmental fluctuations set at a temperature change of ≥±3°C / 10 minutes and a CO2 concentration change of ≥200 ppm / 10 minutes. During the model training phase, K-Means was used for preliminary unsupervised clustering to calibrate the feature distributions under different operating states. State annotation was then performed through manual labeling and supervised learning. The building status analysis module updates the current operating status of the building at a minute-by-minute frequency and predicts the operating trend for the next 30 minutes (based on an LSTM prediction model), providing a basis for dynamic energy-saving control. For example, during peak office hours, the system detected a sharp rise in CO2 accompanied by an increase in occupancy density, immediately identifying a "high-load operation" state and increasing fan speed to enhance ventilation efficiency. After obtaining complete data on internal environmental parameters and occupancy density, combined with building operating status information, a multimodal fusion model was used to achieve comprehensive perception of the building environment. This fusion uses a multimodal fusion network based on the Transformer architecture to embed features and model interactions between data from different sensor channels (environment, image, and status). Each modality is first encoded into a unified tensor format, for example, the dimension of the environmental parameter vector is (N, 4), the density of people in the video is (N, 1), and the operating status of the building is (N, 1), where N is the time dimension. Through the multi-head attention mechanism, the model captures the dependencies between the modalities, especially the " The implicit correlation between "sudden concentration increases" and "crowds" was revealed. In the experimental setup, the sampling frequency was once every 5 seconds, and 30 days of data were continuously trained. The fusion accuracy was evaluated through time series reconstruction error, and the final fusion accuracy reached 92.3%. The output multimodal comprehensive perception data includes the perception status label of each area of ​​the building (such as "intensive ventilation", "comfortable and stable", and "crowded and stranded") and the perception confidence, providing a unified input interface for subsequent dynamic scene analysis.

[0028] Multimodal integrated sensory data is incorporated into the spatiotemporal modeling phase. A spatiotemporal encoding model based on a graph neural network (GNN) is constructed to extract cyclical and anomalous patterns in building usage behavior. The building space is divided into several sub-areas, each of which is considered a node in the graph. Node states in different time periods are connected by edge weights, forming a spatiotemporal graph structure. A spatio-temporal graph convolutional network (ST-GCN) is used to convolutionally extract node features. Positional encoding and a sliding time window mechanism (with a window length of 1 hour and a step size of 15 minutes) are also introduced to enhance the identification of cyclical behavior. For example, a typical "CO2 rise and dense crowds between 9:00 and 11:00 AM" pattern detected in office areas is considered a high-frequency behavior pattern. A dense crowd gathering at 3:00 AM on a particular day is classified as an anomalous behavior pattern. The model was trained on 60 consecutive days of multimodal data, ultimately extracting 18 typical building behavior patterns and defining their corresponding behavior description templates (such as "centralized meetings," "after-hours work," and "short-term high-temperature stays") to provide behavioral basis for energy-saving strategy development. Finally, building behavior patterns were mapped to dynamic operating scenarios, and a systematic scenario evolution network was constructed using a dynamic scenario ontology graph. Graph nodes represent specific behavioral scenarios (e.g., "normal office work," "lunch break," and "abnormal gathering"), while edges represent scenario transitions. Graph construction relies on behavioral pattern sequences and their temporal transition probabilities, combined with environmental parameter distributions to model scenario evolution. To explore the spatial variations of environmental parameters, high-dimensional interpolation methods (such as Kriging interpolation) were used to construct spatial heat maps of carbon dioxide concentration and temperature and humidity, and their temporal evolution was analyzed using a convolutional temporal network (TCN). Each edge in the graph is accompanied by parameter change indicators (e.g., temperature rise rate, carbon dioxide concentration diffusion radius), indicating environmental change trends. Experimental results show that the constructed graph can effectively predict scenario transitions in the short term (with an average lead time of 12 minutes) and identify high-energy consumption triggering nodes. By integrating with fan variable frequency control systems, air volume can be adjusted in advance to optimize energy consumption distribution. For example, the fan speed is automatically increased by 10% 30 minutes before the predicted "personnel peak + ventilation lag" scenario, thereby reducing the carbon dioxide retention time during the peak period by 22% and improving energy saving efficiency by about 8.3%.

[0029] In this embodiment, refer to Figure 3 , is a flowchart of the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Obtaining historical human operation logs of centrifugal fans; performing abnormal data detection on the human operation logs, and performing adaptive abnormality filtering processing to obtain abnormality filtering optimization logs; Identify operational behaviors based on the abnormal filtering and optimization logs, and extract sequences of human operation events; Based on the sequence of human operation events, we analyze the work and rest patterns of employees, air conditioning usage habits, manual intervention frequency, and energy-saving preference levels to build a user-device interaction behavior model. User usage preferences are mined in the user-device interaction behavior model, and device response matching is performed to build a personalized wind turbine usage behavior portrait.

[0030] In this embodiment, historical operation logs need to be collected from the backend of the wind turbine control system. Log sources primarily include the building automation system (BAS), device control panel interfaces (such as Modbus and BACnet protocols), and manual operation terminals (such as touchscreen HMIs and smart building control apps). Operation logs include multiple fields, such as timestamp, operator ID, operation type (e.g., on, off, frequency adjustment), operation value (e.g., frequency adjustment from 45Hz to 60Hz), and control mode (manual / automatic). Log data is stored in a structured format in the device operation database, with a typical log collection cycle of 1–10 seconds. To ensure the accuracy of subsequent analysis, anomaly detection and adaptive filtering are required for log data. A hybrid algorithm combining an isolation forest and a sliding window-based time series mutation detection method is used to identify anomalies in log data. The isolation forest method is suitable for detecting isolated operation records that deviate from behavioral patterns, such as frequent on / off operations late at night or high-frequency fluctuations. The sliding window method (with a 30-minute window size and a 10-minute step size) is used to identify discrete noise or system false alarms within a continuous time period. In a practical experiment, six months of wind turbine operation logs from a commercial building were examined, totaling over 400,000 operation records. Approximately 2.8% of these records contained abnormalities, including batch misoperations (multiple on / off cycles) and unexpected frequency conversion adjustments (frequency jumps exceeding 15Hz). These abnormal logs were filtered through context reconstruction, time period compliance verification, and data correction rules. The resulting optimized logs provide a clean, high-confidence data foundation for subsequent user behavior modeling. Based on the acquired and cleaned optimized logs, the operational behavior recognition phase begins. The goal is to extract structured "human-operated event sequences" from the lengthy log data. First, the log data is time-aligned and windowed, with each day divided into 96 15-minute time periods to capture high-frequency and regular operational behaviors. A hybrid rule-based and learning-based event recognition approach is employed. By setting key trigger thresholds (e.g., frequency changes exceeding 10Hz or more consecutive on / off cycles), cluster analysis is performed on the logs to identify operational intent. Machine learning classification models (e.g., Support Vector Machines (SVMs) and XGBoost) are then used to confirm the operation type, preventing the misidentification of automated system operations as human actions. Event sequences are represented as triples (operator ID, operation type, and operation time), forming a structured dataset suitable for modeling. In the experiment, a one-year log collected from a municipal office building was analyzed, extracting 14,260 valid human operation events, over 70% of which were manual activation or frequency adjustment operations. Further analysis revealed clear time periods for human operation, such as frequent fan activation during the morning rush hour (7:30–9:00 AM) and fine-tuning of temperature and frequency during the late lunch break (1:00–2:00 PM).This structured event sequence not only clearly depicts the daily human control characteristics of the wind turbine, but also provides a standardized behavioral data basis for subsequent personnel habit analysis and model training.

[0031] Event sequence data allows for in-depth analysis of users' work and rest patterns, usage habits, and energy-saving preferences. This phase employs time series clustering analysis and frequent pattern mining to identify user behavior patterns. First, the operation event sequences are aggregated at the user level, and K-Means and DBSCAN are used to segment users, identifying groups with similar behavior patterns. For example, some users may switch on and off at similar times daily, indicating a clear work and rest schedule; while others may have scattered operations, potentially indicating random adjustments or temporary interventions. Frequent item set algorithms (Apriori or FP-Growth) are used to extract high-frequency operation patterns. For example, "turn on the fan from 07:50–08:10 daily and increase the frequency to 60Hz at 08:15" is a typical behavior chain. High-density operation periods and behavioral continuity are then identified through visualization methods such as heat maps and behavior trajectory diagrams. Energy-saving preference indicators are defined by combining operation frequency and operation magnitude (such as average frequency adjustment magnitude and on-time duration). For example, "users with a high energy-saving preference" typically only turn on equipment for short periods of time when necessary and tend to operate at a low frequency; whereas "users with a low energy-saving preference" often run fans at a high frequency for long periods of time. In actual building scenarios, over 80% of users tend to have a moderate preference level, and less than 5% engage in extreme behavior. However, their operating behaviors have a significant impact on overall energy consumption (which can exceed average energy consumption by more than 15%). After obtaining the user's operating behavior sequence and their energy-saving preference level, it is necessary to further construct an interactive behavior model between the user and the fan equipment. This model is based on a modeling framework that combines Markov decision processes (MDPs) with Bayesian behavior graphs. It can describe the dynamic relationship between user operating intentions, behavior triggering probabilities, and device response states.

[0032] The model defines a state space (fan operating status, time period, operational context, etc.), an action space (on, off, variable frequency increase / decrease, etc.), and transition probabilities. For users in an administrative building, the model defines the state "starting during the morning rush hour on weekdays." The subsequent behavior path is primarily "manual start + frequency increase," with a probability of approximately 0.86. However, the probability drops to 0.12 for operations during the same time period on non-weekdays. A Bayesian graph structure is also introduced to model user behavioral intentions, inferring current operational intentions (e.g., "adjusting comfort" or "saving energy and reducing consumption") from multiple observed variables (e.g., time period, previous operation type, and CO2 concentration). The model adjusts its parameters in real time based on sensor status, user behavior, and system feedback, enabling adaptive behavior prediction. After completing user behavior modeling, the final step is to deeply explore user preferences and construct a personalized fan usage profile. This profile is composed of four core dimensions: work-life rhythm intensity, operation frequency level, energy-saving preference score, and manual intervention dependency. A weighted aggregation approach is used to form the user profile vector. Preference mining utilizes a hybrid recommendation mechanism based on factorization machines and collaborative filtering to analyze the interactive relationship between user history and environmental conditions. For example, when CO2 concentrations rise above 1000 ppm, 60% of users proactively increase the frequency, while some show no significant response, indicating insensitivity or laziness to adjust, thus defining them as "low-intervention" users. Integrating user profiles, the system generates personalized control recommendation templates for each user, such as "recommending automatic control mode + flexible frequency adjustment" or "prioritizing energy-saving control scenarios + behavior-predictive linkage control." Spectral profiles (such as radar graphs) visualize users' positions within a multi-dimensional behavioral space, enabling hierarchical management of building user profiles. Deep integration with wind turbine control systems enables personalized adaptation of control strategies. For example, if a user with a "high-frequency intervention + high energy consumption preference" is detected entering the control interface, the system automatically limits the maximum frequency and provides energy-saving recommendations. In actual deployments, this personalized control response model can reduce unnecessary energy consumption by over 10%, improving both user satisfaction and comfort.

[0033] In this embodiment, reference Figure 4 The above is a flowchart of the detailed implementation steps of step S3. In this embodiment, the detailed implementation steps of step S3 include: Obtain the current weather forecast information stream; perform weather trend analysis on the current weather forecast information stream and extract the weather time series change trend curve; Predict the potential environmental parameter changes in the building based on the weather time series trend curve and generate the predicted value of the potential environmental parameter changes; Based on the predicted value of potential environmental parameter changes, the scene state changes are evolved and the scene evolution situation characteristics are constructed; Based on the dynamic scene operation map and personalized fan usage behavior portrait, the usage pattern of fans in buildings is mined to obtain the usage pattern of fans in buildings; Based on the scene evolution characteristics and the usage patterns of fans in the building, the fan operation demand is predicted to obtain the fan operation demand forecast value; The fan load demand is calculated based on the fan operation demand forecast value to generate a real-time fan load demand graph.

[0034] In this example, current and future weather forecast information for the building area is obtained from authoritative meteorological service platforms (such as the China Meteorological Administration API, Weather.com, and OpenWeatherMap). The information stream covers hourly weather data for at least the next 72 hours, primarily including parameters such as temperature, relative humidity, wind speed and direction, atmospheric pressure, precipitation probability, UV index, and cloud cover. In the experimental deployment, the system updates weather information hourly through scheduled API calls. The collected meteorological data is transmitted to the edge server via middleware (such as Kafka) and converted to a standard format (JSON or CSV structured form). Each weather data record carries a timestamp, geographic coordinates, and multi-dimensional meteorological parameters. For example, a weather stream collected from an office building on a certain day in July 2025 included a 24-hour temperature range from 29°C to 37°C, a relative humidity drop from 48% to 26%, and intermittent southerly winds of force 2–4. After acquiring the weather information stream, time series analysis is performed to extract key trend characteristic curves and form a weather evolution map. First, meteorological data is normalized using a sliding window (window length 3 hours, step length 1 hour) to smooth out sudden changes and enhance trend identification. Time series decomposition methods, such as STL decomposition (Seasonal-Trend decomposition using Loess), are then used to decompose each meteorological variable into long-term trends, cyclical fluctuations, and short-term residuals. For variables that primarily influence the building's internal environment (such as temperature, humidity, and wind speed), a dual-model forecasting mechanism based on ARIMA and LSTM is further introduced to perform extended forecasts and fits for meteorological trends over the next 24 hours, outputting complete trend curves. For example, during testing in mid-July 2025, the system's temperature trend forecasts demonstrated an accuracy of 93.1%, enabling early identification of daytime peak high temperature periods (e.g., 2:00 PM–4:00 PM) and critical nighttime temperature drops. The trend curves generated by the system are plotted with time on the horizontal axis and meteorological element changes on the vertical axis, forming a data structure with time-series evolutionary characteristics. These trend curves are not only used to predict environmental parameters but also serve as drivers of dynamic scene evolution, assisting in identifying proactive trends in building equipment adjustment needs.

[0035] After obtaining the weather trend curve, its impact on the building's internal microenvironment needs to be further deduced to form a prediction of potential environmental parameter changes. This prediction model is constructed based on a coupled building thermal and humidity response model and a meteorologically driven model, primarily considering the following core influencing pathways: outdoor temperature and humidity → heat conduction through the building envelope → dynamic changes in indoor temperature and humidity. Using a gray-box modeling approach, parameters such as building heat capacity, building envelope thermal conductivity, window-to-wall ratio, and ventilation rate are modeled and calibrated using historical environmental monitoring data. Input variables include predicted external temperature, humidity, and wind speed, and the output is predicted values ​​for environmental parameters such as indoor temperature, humidity, and carbon dioxide concentration. For example, in one experimental building, the building simulation model parameters were set as follows: overall heat capacity of 1700 kJ / K, window-to-wall ratio of 0.3, and ventilation efficiency of 2.0 times / hour. The average deviation between the model predictions and the measured environmental changes did not exceed ±0.9°C.

[0036] An LSTM (Long Short-Term Memory) network is introduced to perform temporal learning correction on the residual error, improving the model's adaptability to nonlinear responses caused by short-term disturbances (such as gusty winds or sudden rainfall). The final prediction output is a potential environmental change value with a 15-minute resolution for the next six hours, providing critical support for subsequent scenario evolution assessment and device control strategy formulation. Based on the predicted indoor environmental change values, combined with the current building space state (such as occupancy density, area access status, etc.) and behavioral pattern maps, the system constructs a "sequence of possible future operational scenarios" for the building. This process utilizes a scenario evolution model based on a state transition probability matrix to infer and classify the building's scenarios over several future time periods. Each scenario state is represented by a four-tuple: {area ID, occupancy density level, environmental load level, operational mode (automatic / intervention)}. By introducing a time-stepping propagation mechanism, the current state is input into the state transition engine, which infers the next state based on weather trends and environmental forecasts, and records the state transition path. For example: the current state is "medium load + medium density + afternoon period". If the external temperature rises by 5°C and the population density rises to 1.2 people / ㎡, the system simulation state will enter the "high load + high density" scenario.

[0037] During the continuous simulation process, the system visualizes scenario evolution trends as multiple time-dependent state trajectories and extracts key turning points, load peaks, and high-frequency state segments to form a scenario state feature matrix. In experiments, modeling was performed for three building types: office buildings, teaching buildings, and libraries. The scenario evolution prediction accuracy ranged from 88% to 92%, providing advance data support for active fan control. After establishing the dynamic scenario evolution, the system, combined with the constructed personalized fan usage behavior profiles, entered the in-depth exploration phase of fan usage patterns. The goal of this phase is to model the causal relationships between different scenario states and user control behaviors, extract regular fan usage patterns, and optimize predictive operation strategies. A combination of decision trees and Bayesian causal networks is used to model user response behaviors: scenario state serves as the input variable, user profile vectors serve as the control variables, and fan operation behaviors (frequency adjustment, start / stop control) serve as the response variables. The model is trained based on over 1,000 hours of historical operation records and scenario state mapping data, extracting fan usage preference paths for different user categories under different scenario states. For example, for users with a "high intervention and high energy-saving preference," under "high temperature and medium density" conditions, over 80% choose to adjust the frequency to 45Hz for less than 30 minutes. Meanwhile, users with a "low intervention and comfort priority" preference prefer long-term high-frequency operation within the 55–60Hz range. These patterns are refined into a rule base and usage policies, forming a user- and scenario-driven usage pattern model that provides a direct behavioral driver for demand forecasting. Based on scenario dynamics and the fan usage pattern model, the system can predict fan operation demand within a specific future timeframe. This prediction process utilizes a hybrid neural network architecture (RNN + attention mechanism) integrated with a rule engine. It takes the future scenario state sequence as the temporal input sequence and combines user behavioral intentions with predicted environmental changes to output a prediction of the fan operating state at each time step. Each prediction output consists of two core values: the expected operating state (on / off) and the target frequency range (e.g., 45–60Hz), along with an accompanying confidence score. In the experimental deployment, training and testing of operational demand over a 48-hour rolling forecast cycle achieved a prediction accuracy of 89.7%, demonstrating strong sensitivity and foresight in office areas with high-frequency demand fluctuations. The acquisition of wind turbine operational demand forecasts provides a quantitative input foundation for the wind turbine load scheduling system and lays a technical foundation for proactive energy-saving control.

[0038] Based on the predicted fan operating demand values, combined with the building area division, equipment specifications and inverter response curve, the fan load demand in each time slice is calculated and visualized as a "real-time fan load demand graph". This graph shows the load distribution of each fan in the building area at different time periods, including key parameters such as target frequency, expected operating time, and load percentage. The load calculation model is based on the fan performance curve (P = k × ) and establishes a relationship between air volume and frequency, factoring in correction factors for environmental factors (air density, filtration resistance) to form precise load estimation logic. The system outputs a time series matrix with a 15-minute unit time slice, achieving regional-level fan instantiation granularity. For example, in actual office building testing, the generated 24-hour fan load demand graph showed peak fan load from 9:00 AM to 11:30 AM, with an average frequency of 58 Hz and 87% of maximum power. Loads then increased again from 2:00 PM to 3:30 PM due to high temperatures and a high concentration of people. This graph also serves as a core basis for energy consumption monitoring, energy-saving assessments, and equipment scheduling strategy optimization.

[0039] In this embodiment, the specific steps of mining the usage patterns of fans in a building based on the dynamic scene operation graph and the personalized fan usage behavior portrait to obtain the usage patterns of fans in the building are as follows: Identify sudden changes in the fan's operating status based on personalized fan usage behavior profiles and mark sudden changes in the fan's operating status; Calculate the timestamp of the sudden change point of the wind turbine operating status; Perform periodic mutation distribution according to the timestamp to obtain a periodic mutation distribution graph; Mining the operation trigger factors of the sudden change points of the wind turbine operation state to obtain the human operation trigger factors under the sudden change of state; Predicting the trigger conditions before and after based on the human operation trigger factors to generate the trigger prediction conditions before and after; Based on the dynamic scene operation map, deep learning of the fan usage rules is performed on the pre- and post-trigger prediction conditions to obtain the fan usage rules in the building.

[0040] In this embodiment, identifying sudden changes in operating status from continuous wind turbine operating data requires combining existing user profile models with actual operating trajectories to determine which operational changes are "unexpected" or "rapid and drastic fluctuations." Sudden changes in operating status typically manifest as dramatic changes in wind turbine frequency (e.g., from 40Hz to 60Hz), sudden switching between start and stop states, and discontinuous fluctuations in power / air volume. First, the system collects real-time wind turbine operating data from the building equipment control platform, including operating frequency, start and stop signals, output power, and load current, and stores them in time series. A sliding window-based differential detection algorithm (with a window width set to 5 minutes) is then introduced to calculate a "rate of change index" using continuous first-order differences. When this rate exceeds a predefined sudden change threshold (e.g., Δf > 10Hz / 5 minutes or a power sudden change exceeding 20%), the system preliminarily identifies an operational sudden change point. However, to avoid misjudging adjustments caused by system automatic policy changes or environmental load interventions, the system also utilizes the user's personalized usage behavior profile to analyze whether the time period falls within the user's normal control cycle, frequently used frequency range, and customary start and stop patterns. For example, if a user profile indicates a preference for starting the fan around 8:00 AM daily and setting the frequency to 55 Hz, and the current mutation point occurs at 10:30 PM and is accompanied by a sudden increase in frequency, this is more likely to be identified as an abnormal mutation. The system further filters out possible strategic changes using the "profile deviation index," retaining only those mutation behaviors that do not conform to the profile control logic and clearly marking them. This ultimately forms a "fan operation mutation point event library," providing foundational data for subsequent analysis. After identifying the fan mutation points, the system must precisely time-locate each mutation behavior to construct a temporal evolution structure and periodicity model. Timestamp extraction relies on a high-precision time synchronization mechanism. All operating data collection sources (including the fan variable frequency controller, BAS system, and IoT sensor nodes) must be synchronized using the NTP protocol to ensure millisecond-level time consistency. The mutation point timestamp extraction process utilizes a sliding window fastest change location algorithm: a 5-minute sliding window is used, and the midpoint of the maximum difference is taken as the mutation occurrence time. For example, when a wind turbine's frequency increases from 38Hz to 60Hz between 07:50 and 07:55, the system identifies the maximum mutation rate point at 07:52:13, which is the mutation timestamp. All timestamp data is recorded in a triplet format: {wind turbine number, mutation type, timestamp}, and uniformly converted to UTC format to facilitate synchronized analysis across multiple devices. In the experiment, one month of data from 24 wind turbines in five buildings was analyzed, extracting a total of 983 high-confidence mutation points with a time accuracy of within ±1 second. Timestamps are key anchor points for constructing subsequent periodic distributions, causal path analysis, and deep learning modeling. The system also requires alignment and tagging with external events (such as access control records and weather change times) to achieve event-level time synchronization, laying a solid foundation for the accuracy of subsequent periodic modeling and prediction.

[0041] The goal of periodic mutation distribution modeling is to identify the temporal clustering and regularity of wind turbine operational mutations. By mapping mutation timestamps to standard daily (0:00–23:59) and weekly (Monday–Sunday) dimensions, the statistical distribution of these events across various time periods is analyzed to determine whether mutation events are associated with typical operating behavior cycles, usage scenarios, and building schedules. First, timestamps are normalized to "intraday time points," and a 96-slot distribution vector is constructed with 15-minute units. All weekly mutation events are then mapped to these time slots. A weekly periodic distribution table is also constructed, annotating events with labels such as weekday / weekend, daytime / nighttime, etc. Cluster analysis is then performed on the mutation frequencies, using K-means to categorize time periods. For example, clustering is performed during the morning peak (7:00–9:00), midday (12:00–14:00), and evening (18:00–20:00) as cluster centers for group identification. Finally, a "periodic mutation distribution map" is output as a heat map, displaying the frequency of mutation events within each time period or weekday. System analysis results show that sudden changes often occur during periods of high manual operation (such as start-up during the workday, afternoon adjustments, and shutdown just before the end of the shift). Weekend sudden changes significantly decrease, accounting for only 18.4% of the total. This periodic distribution not only reveals the behavioral rhythm of sudden changes in wind turbine operation but also provides a time window screening logic for subsequent backtracking of triggering factors, helping to develop more precise timing rules for control strategies. After determining the periodic distribution, the system further traces multidimensional data around the sudden change point to identify the human factors that may have triggered the sudden change in wind turbine behavior. This process relies on the fusion of multi-source information, including user operation logs (such as app control and physical button records), environmental sensor data (CO2 concentration, temperature and humidity changes), changes in occupancy density from video analysis, and even external events (such as sudden weather changes and power outage restoration). To accurately identify trigger paths, the system extends the analysis window forward and backward by 10 minutes, centered on the sudden change point timestamp. All key variables within the window are extracted and uniformly encoded into a time series feature matrix. A decision tree method combining Apriori frequent pattern analysis is used to identify highly correlated trigger factor combinations. For example, if a wind turbine experienced a remote frequency adjustment via an app five minutes before a sudden change, CO2 concentrations exceeded 1200 ppm, and the number of people in the area exceeded 15, the system identified this combination of variables as a high-confidence trigger. In measured data, approximately 76% of sudden changes can be clearly traced back to manual operation records, primarily due to app control, accounting for 51.3%. Other triggers include local HMI operation (22%) and manual intervention triggered by environmental parameter thresholds (15%). The system encodes these causal chains and outputs them as a "human operation trigger rule library," providing prior knowledge support for behavior prediction and control strategy formulation.

[0042] After identifying key triggers, the system models them to predict whether similar combinations of conditions could cause a sudden change in the wind turbine's operating state in the future, thereby constructing a "pre- and post-trigger prediction condition" system. The model inputs include building status characteristics within the current time period, including environmental parameters (temperature, humidity, air quality), occupancy density, current wind turbine status (frequency, load), user activity (online status, last operation time), and historical frequency of sudden changes. The output is the likelihood of a sudden change in the next time period and its type (start / stop / frequency transition / operating mode switch). To enhance predictive capabilities, the system utilizes a dual-model architecture: random forest and LSTM. The former processes nonlinear combinations of structured feature variables, while the latter processes time series input variables. Each prediction result is accompanied by a confidence interval (e.g., a 72% probability of sudden change expected within 15 minutes) and a categorical labeling of the prediction type. In experimental data, the model achieved an overall prediction accuracy of 89.2% on the test set, with prediction accuracy exceeding 93% for high-frequency users. The system ultimately generates a list of pre- and post-trigger prediction conditions, centered around the formula "condition set → trigger probability → predicted response." This provides real-time input for generating dynamic fan control strategies and is also used for intelligent early warning, policy fine-tuning, and behavioral simulation. These prediction conditions, along with prior trigger factors and periodic mutation behaviors, are embedded into a dynamic scenario operation graph. Global usage patterns are modeled using a graph-based deep learning approach (e.g., a GNN + a temporal transformer). Each scenario node contains the current environmental state (temperature, humidity, CO2), occupant state (density, activity index), operational state (fan status, load level), user state (preference level, intervention tendency), and a prediction condition label. The system performs multi-hop correlation at the path level to identify which pre-condition evolutions are most likely to trigger a sudden response and which user profile characteristics are critical to the fan's response pattern. For example, in a scenario characterized by repeated occurrences of "high midday load, high-intervention users, and open windows," a sudden increase in fan frequency is flagged as a high-probability event. The model then learns this pattern as a "strong feedback operation path" and automatically adjusts the frequency in anticipation of similar scenarios in the future. The system uses an attention mechanism to dynamically focus on mutation nodes and prediction factors, enabling robust modeling in irregular scene switching. In an experiment deployed in a research building, the system constructed 17 high-frequency path maps within two weeks, covering approximately 83% of mutation event types, with an average prediction accuracy of 91.7%. This system has become a core technical support for optimizing wind turbine cluster coordinated control, energy consumption balance, and improving user experience.

[0043] In this embodiment, step S4 includes the following steps: Collect real-time operating parameters of centrifugal fans, including current, voltage, frequency, speed, temperature change curves, and vibration signal spectrum characteristics, and perform fitting to obtain a heterogeneous operating data set; Perform multi-scale time-frequency decomposition on heterogeneous operating data sets to extract wind turbine operating characteristics at multiple levels; Identify the state transition points and key operation nodes of the wind turbine based on the wind turbine operation characteristics, and extract the key operation nodes and state transition points; Based on key operating nodes and state transition points, implicit correlation analysis between parameters is performed, and high-order feature space mapping is performed to construct a multi-dimensional wind turbine operation vector space; Principal component dimensionality reduction analysis and adaptive feature coding are performed on the multi-dimensional wind turbine operation vector space to construct the wind turbine real-time operation condition matrix.

[0044] In this embodiment, collecting real-time operating parameters of the centrifugal fan forms the data foundation for the entire system. Key monitored parameters include current, voltage, frequency, speed, body temperature curves, and vibration signal spectral characteristics. Data collection is typically accomplished using high-precision sensors installed in key locations on the fan. The sampling frequency for current and voltage signals is typically set above 1kHz to capture rapidly changing characteristics. Frequency and speed data are read in real time by the fan controller, while temperature is sampled at 1-second intervals using thermocouples or infrared sensors. Vibration signal acquisition is particularly critical. An accelerometer is used to collect casing vibration data, typically sampling at a frequency of 10kHz to ensure complete spectral information. The collected heterogeneous data is synchronized and aligned on the time axis, employing timestamp alignment technology to ensure a consistent time base for all data. The collected data is then preprocessed, including denoising, missing value imputation, and normalization, to ensure that it meets the requirements of subsequent analysis. To construct a heterogeneous operating dataset, the system fuses these multi-source data into a unified structured format, forming a comprehensive dataset containing multi-dimensional information such as time-series electrical signals, mechanical speed and temperature waveforms, and vibration spectra. This facilitates comprehensive analysis of the fan's operating status from multiple angles. A multi-scale time-frequency decomposition of the constructed heterogeneous operating dataset was performed to extract rich wind turbine operational characteristics. This step utilizes a combination of Discrete Wavelet Transform (DWT) and Short-Time Fourier Transform (STFT) to obtain both temporal localization information and the dynamic evolution of frequency components. First, the vibration signal is decomposed using multi-layer wavelet decomposition, typically into four layers, to obtain energy characteristics in different frequency bands, from low to high. This reveals the different possible fault frequency bands of the wind turbine's mechanical components. For time series signals such as current, voltage, frequency, and speed, an STFT analysis was performed with a window length of 256 points and a 75% overlap rate to extract a two-dimensional time-frequency spectrum, further refining the wind turbine's electrical characteristics and mechanical synchronization. The multi-scale decomposition features, including energy, entropy, spectral centroid, kurtosis, and RMS value, comprehensively reflect the vibration intensity, waveform complexity, and frequency variations of the wind turbine during operation. In experiments, using vibration signals as an example, the system successfully distinguished between normal operation and slightly loose bearings using the energy characteristics of different frequency bands after wavelet packet decomposition, achieving a classification accuracy of 92%. Through this multi-scale time-frequency domain decomposition, the system achieves fine-grained perception of wind turbine operating conditions, providing a rich feature foundation for subsequent key node identification and status assessment.

[0045] State transition points are critical turning points in wind turbine operation, such as startup, shutdown, sudden load changes, or abnormal vibration events. Critical nodes, on the other hand, are time points that significantly impact performance and safety throughout the entire operation process. This step utilizes a change point detection algorithm, combined with cumulative statistics and control chart methods, to detect anomalies in feature sequences and identify state transitions. Specifically, the mean and variance changes within a sliding window are calculated for the multi-scale feature time series. A threshold is then set to identify change points. The typical window length is 60 seconds, and the threshold is trained based on historical data. For example, if the vibration spectrum energy jumps from a stable state of 0.15 g^2 to 0.35 g^2 and remains there for more than two minutes, the system marks it as an abnormal transition point. Electrical characteristics, such as a sudden fluctuation of more than 10% in current frequency, are also combined to identify this point as a critical operational node. This method enables the system to efficiently capture state transitions during wind turbine operation. Experimental validation on multiple units demonstrates an accuracy rate of over 90% in identifying state transition points, with a false negative rate of less than 5%, ensuring timely intelligent diagnosis and optimized control. Based on key operating nodes and state transition points, implicit correlation analysis between parameters was conducted, and high-level feature space mapping was implemented to construct a multidimensional wind turbine operating vector space. This correlation analysis employed mainstream correlation coefficient analysis (Pearson and Spearman) and information entropy correlation analysis, identifying implicit coupling relationships between temperature and vibration energy, and current and speed. Furthermore, nonlinear mapping methods, such as kernel principal component analysis (Kernel PCA) and multidimensional scaling (MDS), were applied to project high-dimensional features into a higher-level space, enhancing the expression of nonlinear correlations between features. This mapping not only preserves the diversity of the original features but also allows potential operating modes and abnormal states to be represented as more recognizable vector features. During the experimental phase, by constructing an operating feature vector containing over 30 dimensions, a multidimensional characterization of wind turbine status was achieved, retaining over 85% of the data information both before and after dimensionality reduction. This constructed multidimensional operating space provides accurate and rich input for subsequent machine learning models, facilitating more precise state classification and fault prediction.

[0046] Principal component dimensionality reduction analysis and adaptive feature encoding are performed on the multi-dimensional fan operation vector space to construct a real-time fan operating condition matrix. Principal component analysis (PCA) effectively compresses data dimensions by extracting directions with the highest variance, typically reducing the original 30-dimensional features to 5-7 principal components. These principal components retain over 90% of the original data information, significantly reducing the computational burden. For adaptive feature encoding, the system utilizes an autoencoder to perform nonlinear mapping on the reduced-dimensional data to further extract latent factors and enhance feature representation. Batch normalization and dropout techniques are used during training to prevent overfitting and ensure model generalization. The output is a real-time fan operating condition matrix, which has time as rows and principal components or encoded features as columns, forming a real-time, dynamic snapshot of the operating status. In experiments, this matrix was used for subsequent fault diagnosis, state prediction, and energy-saving control simulations. The overall model response time is kept at the millisecond level, making it suitable for real-time industrial applications and providing a solid data foundation and analytical support for intelligent energy-saving optimization of variable-frequency centrifugal fans.

[0047] In this embodiment, step S5 includes the following steps: Based on the real-time fan load demand diagram, the load calculation of different working conditions is performed on the fan real-time operating condition matrix to obtain the real-time load demand under different working conditions; Pre-adjust the fan frequency and speed according to the real-time load demand, and build load demand prediction and control parameters; Perform multi-objective energy-saving constraint analysis on the load demand forecast and control parameters, and perform minimum energy-saving calculation to obtain the minimum energy-saving control parameters.

[0048] In this embodiment, a real-time wind turbine load demand map is used as a dynamic input. This map, generated by the wind turbine operating demand forecast and load demand calculation in the previous steps, reflects the load distribution characteristics of the wind turbine under different time periods and operating scenarios. The real-time operating condition matrix contains multidimensional operating status information of the wind turbine, both current and historical. By establishing a load demand mapping model, the principal component characteristics in the operating condition matrix are matched with the load distribution in the load demand map. Regression analysis methods (such as multivariate linear regression, support vector regression (SVR), or random forest regression) are typically used to quantify the load demand corresponding to different operating condition characteristics. The model is trained on historical operating data to identify the changing patterns of wind turbine load under various operating condition combinations. In experiments, using one month of real-time data from a factory wind turbine for training, the model achieved a load demand prediction error of less than 5%. This method enables the system to finely distinguish the load demand of the wind turbine under different environments and usage modes, providing accurate input for subsequent adjustments. Based on the load demand forecast results and combined with the dynamic characteristic model of the wind turbine, a real-time frequency and speed adjustment plan is formulated. The specific approach is to first convert the predicted load demand into a corresponding motor speed requirement. The corresponding frequency setpoint is then determined based on the fan's performance curve and the inverter's control characteristics. Control parameters include not only the fan frequency but also acceleration and deceleration times, as well as upper and lower speed limits. This ensures smooth operation of the fan in response to load changes, avoiding mechanical stress and energy consumption spikes. A model predictive control (MPC) algorithm is used to optimize these control parameters. This algorithm solves a constrained optimization problem online, adjusting control inputs in real time to ensure a close match between the fan frequency and speed and the load demand. Experimental tests show that under dynamic load fluctuations, this approach reduces fan response time by approximately 20% and energy consumption by approximately 8%, effectively improving energy savings and system stability. A multi-objective energy-saving constraint analysis is performed on the load demand forecast control parameters, and the minimum energy savings are calculated to determine the minimum energy-saving control parameters. The core of this step is to construct a multi-objective optimization model that comprehensively considers multiple objectives, including energy savings, operational safety, equipment lifespan, and user comfort. During the optimization process, the system uses real-time load demand, frequency, and speed control parameters as decision variables, and sets energy-saving constraints (such as maximum allowable energy consumption and minimum air volume demand) and equipment operation constraints (such as speed fluctuation limit and temperature rise threshold). Through evolutionary algorithms (such as genetic algorithms and particle swarm optimization) or gradient descent methods, the model iteratively searches for the optimal control parameter combination under all constraints to ensure maximum energy saving while maintaining safe and stable operation of the fan. The minimum energy-saving control parameters are ultimately output as specific frequency adjustment amplitudes, speed settings, and response strategies. During the experimental phase, the model was verified under multiple actual operating conditions. The results showed that the average energy saving rate increased by 12% and the equipment failure rate decreased by 15%. This control strategy not only reduces energy consumption but also extends the service life of the fan, reflecting the intelligent optimization goal of balancing energy saving and reliability.

[0049] In this embodiment, step S6 includes the following steps: Execute dynamic fan control based on minimum energy-saving control parameters and collect fan monitoring feedback information; Perform timing response calculation based on fan monitoring feedback information to obtain fan timing response parameters; Predicting the expected energy-saving response based on the minimum energy-saving control parameters to obtain the expected response target; Calculate the target deviation of the wind turbine timing response parameters according to the expected response target to obtain the target difference; Based on the target difference, the minimum energy-saving control parameters are reversely adjusted, and dynamic closed-loop iterative corrections are performed to build an intelligent energy-saving iterative optimization engine.

[0050] In this embodiment, the previously optimized minimum energy-saving control parameters are used as dynamic control inputs for the fan, primarily including frequency setting, speed regulation, and acceleration / deceleration strategies. These control commands are transmitted in real time to the fan drive system via the inverter interface, enabling precise control of the fan's operating status. Simultaneously, various sensors installed on-site continuously collect key operating data, such as current, voltage, speed, body temperature, and vibration signals, forming a monitoring feedback information stream. Data acquisition utilizes a high sampling frequency (1kHz for electrical parameters and 500Hz for mechanical parameters) to ensure real-time performance and integrity. The collected data is transmitted to the control center via the Industrial Internet of Things platform for real-time monitoring and recording. In experimental verification, the real-time control and feedback acquisition system achieved millisecond-level response in an office building fan unit operation scenario, ensuring accurate execution of control commands and timely capturing operational status changes, providing a data foundation for subsequent time-series response analysis. Signal processing techniques are used to analyze the feedback data in both the time and frequency domains, calculating key parameters such as the fan's response time, overshoot, and steady-state error, from the issuance of the control command to the actual stable operating state. For example, after a frequency setting change, the time required for the speed to reach the new target from the initial value (response time) and the difference between the peak speed and the steady-state value (overshoot) are measured. Furthermore, spectral analysis is used to calculate the frequency components and amplitude changes of the vibration response, reflecting the dynamic response characteristics of the machine. This step also involves using sliding window technology to perform segmented statistics on continuous data and extract dynamic response curves. In experiments, the average response time of a wind turbine unit was controlled within 3 seconds, and the overshoot was less than 5%. The response parameters provide intuitive performance indicators for optimized control.

[0051] Using historical data and a wind turbine performance model, combined with minimum energy-saving control parameters, the wind turbine's energy-saving effect and operational response under a given control strategy are predicted. Using machine learning prediction models, such as long short-term memory networks (LSTMs) or gradient boosted tree models (GBDTs), real-time control parameters and historical energy-saving data are trained to produce time-series predictions of energy-saving responses. The prediction output includes metrics such as expected energy consumption reduction, operational efficiency improvement, and potential response delay, forming a desired response target. The model continuously corrects prediction errors by verifying the consistency between historical data and actual responses. In an experimental environment, the prediction model maintained an energy consumption prediction error within ±3% over 30 days of continuous data testing, demonstrating high accuracy and reliability, providing clear energy-saving targets for dynamic control. By comparing actual time-series response parameters with the predicted desired response target, the target difference is calculated, reflecting the deviation between the current control effect and the expected energy-saving performance. Specific calculations include response time difference, energy consumption difference, frequency regulation error, and other metrics, forming a multi-dimensional target difference vector. This difference is used to evaluate the effectiveness of the current control strategy and the potential for energy-saving optimization. The system sets reasonable thresholds to determine whether the difference exceeds the allowable range. If so, it triggers subsequent parameter adjustments. In experiments, the fan control system set a response time difference threshold of 0.5 seconds and an energy consumption deviation threshold of 2%. This effectively identifies deficiencies in control execution and provides a basis for dynamic correction. Based on the target difference, the minimum energy-saving control parameters are reversely adjusted, and dynamic closed-loop iterative corrections are performed to build an intelligent energy-saving iterative optimization engine. This step implements the core closed-loop of intelligent control. By feeding back the target difference, adaptive control algorithms (such as gain-adjusted PID control, reinforcement learning, or gradient-based optimization) are used to adjust control parameters. Specifically, the system adjusts the frequency setting and speed target based on the difference information to achieve refined matching to the fan load. During the iterative process, control parameters are continuously adjusted until the target difference converges within the preset threshold. This closed-loop mechanism has self-learning capabilities, adapting to different environments and usage scenarios to continuously optimize energy efficiency. Experiments show that after multiple rounds of iterative optimization, fan energy consumption is reduced by an average of 10% and response time is improved by 15%, achieving stable and significant energy savings. This intelligent energy-saving iterative optimization engine provides strong support for real-time and efficient energy-saving control of variable-frequency centrifugal fans, taking into account both energy saving and operational reliability.

[0052] In this embodiment, an intelligent energy-saving optimization device for a variable frequency centrifugal fan is provided, which is used to execute the intelligent energy-saving optimization method for a variable frequency centrifugal fan as described above, including: The scenario analysis module is used to collect environmental monitoring parameters in the area, perform dynamic operation scenario analysis, and mine the distribution changes of environmental parameters to build a dynamic scenario operation map; The usage behavior module is used to obtain historical centrifugal fan human operation logs, conduct user usage preference mining and device response matching, and build a personalized fan usage behavior profile; The demand forecasting module is used to mine fan usage patterns within buildings based on dynamic scenario operation graphs and personalized fan usage behavior portraits, and to forecast fan operation demand and generate real-time fan load demand graphs; The fan operating condition analysis module is used to collect the real-time operating parameters of the centrifugal fan, perform principal component dimensionality reduction analysis and adaptive feature coding, and construct the fan real-time operating condition matrix; Energy-saving constraint analysis module, which is used to perform multi-objective energy-saving constraint analysis on the real-time operating condition matrix of the fan based on the real-time fan load demand diagram, and perform minimum energy-saving calculation to obtain the minimum energy-saving control parameters; The energy-saving iterative optimization module is used to perform dynamic control of the fan based on the minimum energy-saving control parameters, and perform dynamic closed-loop iterative correction to build an intelligent energy-saving iterative optimization engine.

[0053] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0054] The foregoing description is intended only to provide specific embodiments of the present invention, which are intended to enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest manner consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent energy-saving optimization method for a variable frequency centrifugal fan, characterized in that: The following steps are involved: Step S1: Collect environmental monitoring parameters in the area, perform dynamic operation scenario analysis, and mine the distribution changes of environmental parameters to build a dynamic scenario operation map; Step S2: Obtain historical centrifugal fan human operation logs, conduct user usage preference mining and device response matching, and build a personalized fan usage behavior profile; Step S3: Based on the dynamic scenario operation map and personalized fan usage behavior portrait, the fan usage pattern in the building is mined, and the fan operation demand is predicted to generate a real-time fan load demand map; Step S4: collecting the real-time operating parameters of the centrifugal fan, performing principal component dimensionality reduction analysis and adaptive feature coding, and constructing a real-time operating condition matrix of the fan; Step S5: performing a multi-objective energy-saving constraint analysis on the real-time operating condition matrix of the fan based on the real-time fan load demand diagram, and performing a minimum energy-saving calculation to obtain the minimum energy-saving control parameters; Step S6: Execute dynamic control of the fan based on the minimum energy-saving control parameters, perform dynamic closed-loop iterative correction, and build an intelligent energy-saving iterative optimization engine.

2. The intelligent energy-saving optimization method for a variable frequency centrifugal fan according to claim 1, characterized in that: The specific steps of step S1 are: Using multiple sensors within the building area to collect environmental monitoring parameters within the area, including temperature, humidity, and carbon dioxide concentration; and obtaining video surveillance images within the building area; Calculate the density of people in the area based on the video surveillance image, and perform building operation status analysis to obtain the building operation status; Performing multimodal environmental comprehensive perception based on the environmental monitoring parameters, the density of people in the area, and the operating status of the building to generate multimodal environmental comprehensive perception data; Perform spatiotemporal encoding on the multimodal environmental integrated perception data and conduct architectural behavior pattern analysis to obtain architectural behavior patterns; Dynamic operation scenarios are analyzed based on building behavior patterns, and changes in environmental parameter distribution are mined to construct dynamic scenario operation maps.

3. The intelligent energy-saving optimization method for a variable frequency centrifugal fan according to claim 1, characterized in that: The specific steps of step S2 are: Obtaining historical human operation logs of centrifugal fans; performing abnormal data detection on the human operation logs, and performing adaptive abnormality filtering processing to obtain abnormality filtering optimization logs; Identify operational behaviors based on the abnormal filtering and optimization logs, and extract sequences of human operation events; Based on the sequence of human operation events, we analyze the work and rest patterns of employees, air conditioning usage habits, manual intervention frequency, and energy-saving preference levels to build a user-device interaction behavior model. User usage preferences are mined in the user-device interaction behavior model, and device response matching is performed to build a personalized wind turbine usage behavior portrait.

4. The intelligent energy-saving optimization method for a variable frequency centrifugal fan according to claim 1, characterized in that: The specific steps of step S3 are: Obtain the current weather forecast information stream; perform weather trend analysis on the current weather forecast information stream and extract the weather time series change trend curve; Predict the potential environmental parameter changes in the building based on the weather time series trend curve and generate the predicted value of the potential environmental parameter changes; Based on the predicted value of potential environmental parameter changes, the scene state changes are evolved and the scene evolution situation characteristics are constructed; Based on the dynamic scene operation map and personalized fan usage behavior portrait, the usage pattern of fans in buildings is mined to obtain the usage pattern of fans in buildings; Based on the scene evolution characteristics and the usage patterns of fans in the building, the fan operation demand is predicted to obtain the fan operation demand forecast value; The fan load demand is calculated based on the fan operation demand forecast value to generate a real-time fan load demand graph.

5. The intelligent energy-saving optimization method for a variable frequency centrifugal fan according to claim 4, characterized in that: The specific steps for mining the usage patterns of fans in buildings based on the dynamic scenario operation graph and the personalized fan usage behavior portrait to obtain the usage patterns of fans in buildings are as follows: Identify sudden changes in the fan's operating status based on personalized fan usage behavior profiles and mark sudden changes in the fan's operating status; Calculate the timestamp of the sudden change point of the wind turbine operating status; Perform periodic mutation distribution according to the timestamp to obtain a periodic mutation distribution graph; Mining the operation trigger factors of the sudden change points of the wind turbine operation state to obtain the human operation trigger factors under the sudden change of state; Predicting the trigger conditions before and after based on the human operation trigger factors to generate the trigger prediction conditions before and after; Based on the dynamic scene operation map, deep learning of the fan usage rules is performed on the pre- and post-trigger prediction conditions to obtain the fan usage rules in the building.

6. The intelligent energy-saving optimization method for a variable frequency centrifugal fan according to claim 1, characterized in that: The specific steps of step S4 are: Collect real-time operating parameters of centrifugal fans, including current, voltage, frequency, speed, temperature change curves, and vibration signal spectrum characteristics, and perform fitting to obtain a heterogeneous operating data set; Perform multi-scale time-frequency decomposition on heterogeneous operating data sets to extract wind turbine operating characteristics at multiple levels; Identify the state transition points and key operation nodes of the wind turbine based on the wind turbine operation characteristics, and extract the key operation nodes and state transition points; Based on key operating nodes and state transition points, implicit correlation analysis between parameters is performed, and high-order feature space mapping is performed to construct a multi-dimensional wind turbine operation vector space; Principal component dimensionality reduction analysis and adaptive feature coding are performed on the multi-dimensional wind turbine operation vector space to construct the wind turbine real-time operation condition matrix.

7. The intelligent energy-saving optimization method for a variable frequency centrifugal fan according to claim 1, characterized in that: The specific steps of step S5 are: Based on the real-time fan load demand diagram, the load calculation of different working conditions is performed on the fan real-time operating condition matrix to obtain the real-time load demand under different working conditions; Pre-adjust the fan frequency and speed according to the real-time load demand, and build load demand prediction and control parameters; Perform multi-objective energy-saving constraint analysis on the load demand forecast and control parameters, and perform minimum energy-saving calculation to obtain the minimum energy-saving control parameters.

8. The intelligent energy-saving optimization method for a variable frequency centrifugal fan according to claim 1, characterized in that: The specific steps of step S6 are: Execute dynamic fan control based on minimum energy-saving control parameters and collect fan monitoring feedback information; Perform timing response calculation based on fan monitoring feedback information to obtain fan timing response parameters; Predicting the expected energy-saving response based on the minimum energy-saving control parameters to obtain the expected response target; Calculate the target deviation of the wind turbine timing response parameters according to the expected response target to obtain the target difference; Based on the target difference, the minimum energy-saving control parameters are reversely adjusted, and dynamic closed-loop iterative corrections are performed to build an intelligent energy-saving iterative optimization engine.

9. An intelligent energy-saving optimization device for a variable frequency centrifugal fan, characterized in that: The method for executing the intelligent energy-saving optimization method for a variable frequency centrifugal fan according to claim 1 comprises: The scenario analysis module is used to collect environmental monitoring parameters in the area, perform dynamic operation scenario analysis, and mine the distribution changes of environmental parameters to build a dynamic scenario operation map; The usage behavior module is used to obtain historical centrifugal fan human operation logs, conduct user usage preference mining and device response matching, and build a personalized fan usage behavior profile; The demand forecasting module is used to mine fan usage patterns within buildings based on dynamic scenario operation graphs and personalized fan usage behavior portraits, and to forecast fan operation demand and generate real-time fan load demand graphs; The fan operating condition analysis module is used to collect the real-time operating parameters of the centrifugal fan, perform principal component dimensionality reduction analysis and adaptive feature coding, and construct the fan real-time operating condition matrix; Energy-saving constraint analysis module, which is used to perform multi-objective energy-saving constraint analysis on the real-time operating condition matrix of the fan based on the real-time fan load demand diagram, and perform minimum energy-saving calculation to obtain the minimum energy-saving control parameters; The energy-saving iterative optimization module is used to perform dynamic control of the fan based on the minimum energy-saving control parameters, and perform dynamic closed-loop iterative correction to build an intelligent energy-saving iterative optimization engine.

Images