A green building fan adaptive speed regulation control method

Abstract

The present application relates to the technical field of adaptive control, in particular to a kind of fan adaptive speed control method for green building.Collect indoor and outdoor environment and operation data, combine building information model data, carry out multi-source heterogeneous data fusion to generate high-dimensional feature vector;Establish dynamic heat balance model, predict future environmental state using LSTM network, generate control reference trajectory;Adaptive control strategy is constructed using DQN network, high-dimensional feature vector and control reference trajectory are input, and initial rotating speed control sequence is generated;Model predictive control framework is constructed, combined with control reference trajectory and initial rotating speed control sequence, rotating speed control sequence is optimized by distributed computing, and optimized rotating speed sequence is generated;By comparing the deviation of closed-loop control prediction state and actual state, adjust LSTM network parameters and DQN strategy and detect abnormality;The present application can improve the effectiveness of adaptive control.

Description

Technical Field

[0001] This invention relates to the field of adaptive control technology, specifically to an adaptive speed control method for fans used in green buildings. Background Technology

[0002] To achieve effective control of complex indoor environmental systems, existing technologies often employ architectures such as Model Reference Adaptive Control (MRAC) or Self-Tuning Regulator (STR). These methods identify the dynamic characteristics of the system online and adjust the controller parameters accordingly to achieve adaptive control. For example, the gain of the PID controller is adjusted by updating the process model parameters online, or the feedforward and feedback control laws in MRAC are adjusted to track the reference model.

[0003] Traditional speed control methods often employ fixed speeds or simple PID control, which struggle to adapt to dynamic changes in the building's internal environment, resulting in poor control performance. While existing adaptive control systems can partially address dynamic adjustment issues, they largely rely on single sensor data and lack multi-source data fusion and global optimization capabilities. Furthermore, current technologies rarely integrate real-time energy consumption targets for green buildings with user comfort requirements, limiting the intelligence level of the control system. Simultaneously, existing control systems lack adaptive learning capabilities and model prediction functions, making it difficult to achieve an optimal balance between energy saving and comfort.

[0004] To address this, an adaptive speed control method for fans used in green buildings is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive speed control method for wind turbines used in green buildings. First, indoor and outdoor environmental and operational data are collected and combined with building information model (BIM) data to perform multi-source heterogeneous data fusion to generate a high-dimensional feature vector. A dynamic thermal balance model is established, and an LSTM network is used to predict future environmental conditions, generating a control reference trajectory. An adaptive control strategy is constructed using a DQN network, taking the high-dimensional feature vector and the control reference trajectory as input to generate an initial speed control sequence. A model predictive control framework is constructed, and the speed control sequence is optimized through distributed computing, combining the control reference trajectory and the initial speed control sequence, to generate an optimized speed sequence. By comparing the deviation between the predicted and actual closed-loop control states, the LSTM network parameters and the DQN strategy are adjusted, and anomalies are detected. This invention can improve the effectiveness of adaptive regulation in IoT control systems.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An adaptive speed control method for fans used in green buildings includes:

[0008] Indoor and outdoor environmental data and fan operation status data are collected through IoT sensor networks. Combined with building information model data, Kalman filtering algorithm is used to fuse multi-source heterogeneous data and generate high-dimensional feature vectors.

[0009] A dynamic thermal balance model of the indoor environment is established, and the high-dimensional feature vector is processed by an LSTM network to predict the future environmental state and generate a control reference trajectory.

[0010] An adaptive control strategy is constructed using a DQN network. The high-dimensional feature vector and the control reference trajectory are used as inputs, and the energy saving rate, PMV thermal comfort index and equipment life are used as multi-objective reward functions to generate an initial speed control sequence.

[0011] A model predictive control framework is constructed, and by combining the control reference trajectory and the initial speed control sequence, the speed control sequence is optimized through distributed computing to generate an optimized speed sequence.

[0012] The optimized rotational speed sequence is input into the dynamic thermal balance model to generate a closed-loop control prediction state. By comparing the deviation between the closed-loop control prediction state and the actual environmental state, the LSTM network parameters and DQN strategy are dynamically adjusted, and the isolated forest algorithm is used to detect anomalies.

[0013] Furthermore, the indoor and outdoor environmental data includes: temperature, humidity, carbon dioxide concentration, and personnel density; the fan operating status data includes: fan speed, fan power, fan vibration, and fan temperature; and the building information model data includes geometric information, material property information, and functional information.

[0014] Furthermore, the process of establishing a dynamic thermal balance model of the indoor environment, processing the high-dimensional feature vectors using an LSTM network, predicting future environmental conditions, and generating a control reference trajectory includes:

[0015] Based on the principles of building thermodynamics, and combined with room volume, material thermal conductivity, and air duct layout data from building information modeling, a dynamic thermal balance model of the indoor environment is established to calculate real-time changes in temperature, humidity, and CO2 concentration.

[0016] The high-dimensional feature vector is input into an LSTM network to process the feature sequence within a preset time window, thereby obtaining the predicted temperature, humidity, and carbon dioxide concentration.

[0017] A dynamic thermal equilibrium model is used to apply physical constraints to the LSTM prediction results, correct the prediction bias, and generate a control reference trajectory in time series format.

[0018] The control reference trajectory is transmitted back from the cloud to the edge device via the MQTT protocol as input for adaptive control and model predictive control.

[0019] Furthermore, an adaptive control strategy is constructed using a DQN network. The process of generating the initial speed control sequence, with the high-dimensional feature vector and the control reference trajectory as inputs, and energy saving rate, PMV thermal comfort index, and equipment lifespan as multi-objective reward functions, includes:

[0020] A DQN network is constructed, defining the discrete action space of the wind turbine speed, with the preprocessed high-dimensional feature vector and the control reference trajectory as inputs;

[0021] Design a multi-objective reward function with energy saving rate, PMV thermal comfort index and equipment life as optimization objectives, calculate reward value, and guide DQN to learn the optimal control strategy;

[0022] DQN is trained in the cloud using a greedy strategy, and the state-action-reward data is stored using an experience replay mechanism to update the neural network parameters.

[0023] DQN inference is performed on the edge device to generate the initial speed control sequence based on the high-dimensional feature vector and the control reference trajectory input in real time.

[0024] Furthermore, a model predictive control framework is constructed. Combining the control reference trajectory and the initial speed control sequence, the speed control sequence is optimized through distributed computing. The process of generating the optimized speed sequence includes:

[0025] A model predictive control framework is constructed, which combines the control reference trajectory and the dynamic thermal balance model, and defines the prediction time domain, control time domain, and state transition function.

[0026] Using the initial speed control sequence as the initial value, an optimization objective function is designed, taking into account the energy saving rate and PMV thermal comfort index, and constraints are set.

[0027] The optimization problem is decomposed into sub-problems based on rooms or air ducts, and solved in parallel on edge devices through distributed computing to generate a local optimized speed sequence. The local optimization results are then aggregated in the cloud to generate an optimized speed sequence.

[0028] Furthermore, the process of dynamically adjusting the LSTM network parameters and DQN strategy by comparing the deviation between the predicted state of the closed-loop control and the actual environmental state, and using the isolated forest algorithm to detect anomalies includes:

[0029] The optimized rotational speed sequence is input into the dynamic thermal balance model to generate the closed-loop control prediction state in time series format;

[0030] Compare the predicted state of the closed-loop control with the actual environmental state, calculate the weighted Euclidean distance as the deviation, and determine whether to trigger a model update.

[0031] Based on the deviation results, the LSTM network parameters and the reward function weights of the DQN strategy are dynamically adjusted in the cloud, the parameters are updated, and the data is sent to the edge devices.

[0032] The isolated forest algorithm is used to analyze the data collected by the IoT sensor network and detect abnormal states. If an anomaly is detected, a backup PID control strategy is triggered.

[0033] Furthermore, this is achieved through a cloud-edge collaborative architecture, where edge devices perform data processing and real-time control, while the cloud is responsible for model training and control optimization.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention proposes a method for generating control reference trajectories. This method combines building thermodynamics (thermal balance model), artificial intelligence (LSTM), and the Internet of Things (MQTT data transmission). The thermal balance model provides physical constraints, and LSTM enables forward-looking environmental prediction to generate control reference trajectories. The control reference trajectories use the predicted environmental state to guide the control and adjustment of fan speed, thereby improving the effectiveness of adaptive speed control of fans used in green buildings.

[0036] 2. This invention proposes a long-term adaptive control method. This method employs a deep Q-network (DQN) to construct an adaptive control strategy, using a high-dimensional feature vector and control reference trajectory as inputs, and employing energy saving rate, PMV thermal comfort index, and equipment lifespan as multi-objective reward functions to generate an initial speed control sequence. The DQN automatically adjusts the control strategy through online learning, reducing reliance on manual parameter tuning, and can handle complex nonlinear environments. It optimizes multi-objective balance through the reward function to achieve intelligent control. This method can improve the effectiveness of adaptive speed control for wind turbines used in green buildings.

[0037] 3. This invention proposes a short-term control optimization method. This method constructs a model predictive control (MPC) framework, combines LSTM predicted trajectory and DQN initial speed sequence, and optimizes the speed sequence through distributed computing. MPC dynamically optimizes the speed according to the real-time environmental state and predicted trajectory to adapt to short-term environmental changes. Distributed computing decomposes the optimization problem into sub-problems to meet the real-time control requirements of green buildings. This method can effectively improve control accuracy, thereby improving the effectiveness of adaptive speed control of fans used in green buildings. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating an adaptive speed control method for a green building fan according to the present invention.

[0039] Figure 2 A schematic diagram illustrating the process of generating the control reference trajectory for this invention;

[0040] Figure 3 This is a schematic diagram illustrating the process of generating the optimized rotational speed sequence for this invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figures 1 to 3 This invention provides an adaptive speed control method for fans used in green buildings, the technical solution of which is as follows:

[0043] Example 1:

[0044] To achieve effective adaptive speed control of fans used in green buildings, a company employed an adaptive speed control method for green building fans proposed in this invention. The flowchart of this method is shown below. Figure 1 As shown, it specifically includes:

[0045] Indoor and outdoor environmental data and fan operation status data are collected through IoT sensor networks. Combined with building information model data, Kalman filtering algorithm is used to fuse multi-source heterogeneous data and generate high-dimensional feature vectors.

[0046] Furthermore, indoor and outdoor environmental data include: temperature, humidity, carbon dioxide (CO2) concentration, and personnel density; fan operating status data includes: fan speed, fan power, fan vibration, and fan temperature; building information model data includes geometric information, material property information, and functional information.

[0047] The following table shows the reference wind turbine operating status data.

[0048] Table 1. Reference Fan Operating Status Data

[0049] Data types Reference values Fan speed 1180rpm Fan power 295W Fan vibration <![CDATA[0.19m / s 2 ]]> Fan temperature 39.7℃

[0050] Furthermore, the geometric information includes the building volume, duct cross-sectional area, duct length, and window area; the material property information includes thermal conductivity and heat capacity.

[0051] Furthermore, the multi-source heterogeneous data fusion process includes: applying the extended Kalman filter algorithm to indoor and outdoor environmental data and fan operating status data, with the state vector defined as [temperature, humidity, CO2 concentration, personnel density, wind speed, fan speed, power], eliminating noise through the state transition matrix and observation matrix, and outputting a smooth data stream; and concatenating the filtered data with the geometric information in the building information model data to generate a high-dimensional feature vector.

[0052] By using indoor and outdoor environmental data, wind turbine operating status data, and building information model data, and by fusing multi-source heterogeneous data, reliable inputs can be provided for subsequent adaptive control and model predictive control, thereby improving the effectiveness of adaptive speed control for wind turbines in green buildings.

[0053] A dynamic thermal balance model of the indoor environment is established, and the high-dimensional feature vector is processed by the LSTM network to predict the future environmental state and generate a control reference trajectory.

[0054] Furthermore, the process of establishing a dynamic thermal balance model of the indoor environment, using an LSTM network to process high-dimensional feature vectors, predicting future environmental conditions, and generating control reference trajectories can be referenced. Figure 2 The details are as follows:

[0055] Based on the principles of building thermodynamics, and combined with room volume, material thermal conductivity, and air duct layout data from building information modeling, a dynamic thermal balance model of the indoor environment is established to calculate real-time changes in temperature, humidity, and CO2 concentration.

[0056] The high-dimensional feature vector is input into the LSTM network, and the feature sequence of the preset time window is processed to obtain the predicted temperature, humidity and carbon dioxide concentration.

[0057] A dynamic thermal balance model of the indoor environment is used to apply physical constraints to the LSTM prediction results, correct the prediction bias, and generate a control reference trajectory in time series format.

[0058] The control reference trajectory is transmitted back from the cloud to the edge device via the MQTT protocol as input for adaptive control and model predictive control.

[0059] Furthermore, the indoor environment dynamic heat balance model includes: temperature heat balance model, humidity balance model and CO2 concentration balance model. These models are all mathematical models based on equations.

[0060] Furthermore, the temperature heat balance model can be expressed as follows: the rate of change of indoor temperature equals the sum of ventilation heat flux, heat dissipation from people, outdoor heat input, and wall heat loss, divided by the product of air heat capacity and room volume; the humidity balance model can be expressed as follows: the rate of change of indoor relative humidity equals the moisture input from ventilation minus moisture loss plus moisture generated by people's respiration, divided by room volume; the CO2 concentration balance model can be expressed as follows: the rate of change of indoor CO2 concentration equals the CO2 generated by people's respiration minus the CO2 removed by ventilation, divided by room volume.

[0061] Furthermore, the LSTM network structure includes: 3 LSTM layers for predicting the future state of the indoor environment, each layer with 128 units, hidden layers using the tanh activation function, a Dropout layer added after each LSTM layer to enhance the model's generalization ability, and a fully connected layer to map the output of the LSTM layer to the predicted value; the preset time window is set to 30 minutes.

[0062] Furthermore, the LSTM prediction results are input into the indoor environment dynamic thermal balance model to check whether energy and mass conservation are satisfied; if they exceed the physical range of the model, linear interpolation is used to correct the predicted values.

[0063] This embodiment combines building thermodynamics (thermal balance model), artificial intelligence (LSTM), and the Internet of Things (MQTT data transmission). The thermal balance model provides physical constraints, LSTM enables forward-looking environmental predictions, and generates control reference trajectories. The control reference trajectories use the predicted environmental conditions to guide the control and adjustment of the fan speed, thereby improving the effectiveness of adaptive speed control of fans in green buildings.

[0064] An adaptive control strategy is constructed using a DQN network. The high-dimensional feature vector and control reference trajectory are used as inputs, and the energy saving rate, PMV thermal comfort index and equipment life are used as multi-objective reward functions to generate the initial speed control sequence.

[0065] Furthermore, an adaptive control strategy is constructed using a DQN network. Taking a high-dimensional feature vector and control reference trajectory as input, and using energy saving rate, PMV thermal comfort index, and equipment lifespan as multi-objective reward functions, the process of generating the initial speed control sequence includes:

[0066] A DQN network is constructed, defining the discrete action space of the wind turbine speed, with the preprocessed high-dimensional feature vector and the control reference trajectory as inputs;

[0067] Design a multi-objective reward function with energy saving rate, PMV thermal comfort index and equipment life as optimization objectives, calculate reward value, and guide DQN to learn the optimal control strategy;

[0068] DQN is trained in the cloud using a greedy strategy, and the state-action-reward data is stored using an experience replay mechanism to update the neural network parameters.

[0069] DQN inference is performed on the edge device to generate an initial speed control sequence based on the high-dimensional feature vector input in real time and the control reference trajectory.

[0070] Furthermore, DQN contains four fully connected neural networks with 128 neurons per layer, and uses the ReLU activation function to enhance the model's nonlinear fitting ability; the discrete action space of the wind turbine speed is divided according to the speed range and step size. For example, for the wind turbine speed range of 0-1800 rpm, the step size is set to 50 rpm, and there are a total of 37 actions.

[0071] Furthermore, the high-dimensional feature vector and the control reference trajectory are normalized before input to improve the model's computational efficiency; the high-dimensional feature vector provides real-time input for DQN; and the control reference trajectory provides the target state for DQN, guiding the calculation of the reward function.

[0072] Furthermore, the multi-objective reward function is represented as a weighted sum of energy-saving reward, thermal comfort reward, equipment lifespan reward, and trajectory tracking reward; the energy-saving reward is represented by 1 minus the ratio of the current fan power to the maximum fan power; the thermal comfort reward is represented by 1 minus the ratio of the absolute value of the current PMV thermal comfort index to the absolute value of the maximum PMV; the equipment lifespan reward is represented by 1 minus the ratio of the current fan vibration intensity to the maximum vibration intensity; and the trajectory tracking reward is represented by 1 minus the ratio of the weighted Euclidean distance between the actual state and the control reference trajectory to the maximum distance threshold.

[0073] Furthermore, the PMV thermal comfort index is calculated using the thermal comfort model formula in the ISO 7730 standard; according to the ISO 7730 standard, the maximum absolute value of PMV is generally set to 3.

[0074] By using DQN online learning to automatically adjust control strategies, the reliance on manual parameter tuning is reduced. At the same time, it can handle complex nonlinear environments and optimize multi-objective balance through reward functions to achieve intelligent control. This can improve the effectiveness of adaptive speed control for fans in green buildings.

[0075] A model predictive control framework is constructed, which combines the control reference trajectory and the initial speed control sequence, and optimizes the speed control sequence through distributed computing to generate an optimized speed sequence.

[0076] Furthermore, a model predictive control framework is constructed. Combining the control reference trajectory and the initial speed control sequence, the speed control sequence is optimized through distributed computing. The process for generating the optimized speed sequence is as follows: Figure 3 As shown, it specifically includes:

[0077] A model predictive control framework is constructed, which combines the control reference trajectory and the dynamic thermal equilibrium model, and defines the prediction time domain, control time domain, and state transition function.

[0078] Using the initial speed control sequence as the initial value, an optimization objective function is designed, taking into account the energy saving rate and PMV thermal comfort index, and constraints are set.

[0079] The optimization problem is decomposed into sub-problems based on rooms or air ducts, and solved in parallel on edge devices through distributed computing to generate a local optimized speed sequence; the local optimization results are then aggregated in the cloud to generate an optimized speed sequence.

[0080] Furthermore, the state transition function of the model predictive control framework adopts the function formula corresponding to the indoor environmental dynamic thermal balance model, namely the temperature thermal balance function, humidity balance function, and CO2 concentration balance function; the state transition function uses the control reference trajectory as the target state; the prediction time domain is 30 minutes, and the control time domain is 5 minutes;

[0081] Furthermore, the optimization objective function of the model predictive control framework calculates the total target value under the prediction time domain steps. The optimization objective function calculates the energy-saving penalty weight multiplied by the square of the current time fan power and the thermal comfort penalty weight multiplied by the square of the deviation between the current time PMV thermal comfort index and the target PMV, and then sums them up. The prediction time domain steps are the ratio of the prediction time domain to the step size, and the step size is 5 seconds.

[0082] Furthermore, the constraints are speed range constraints, PMV range constraints, and power range constraints.

[0083] The initial and optimized speed values ​​at different time steps are given for the initial speed control sequence and the optimized speed sequence, respectively. The initial speed value is obtained by DQN, and the optimized speed value is obtained by Model Predictive Control (MPC). The speed comparison results are shown in Table 2.

[0084] Table 2. Results of Rotational Speed ​​Comparison

[0085] Time step Initial speed value Optimize speed value 300s 1250rpm 1230rpm 900s 1300rpm 1270rpm 1500s 1290rpm 1280rpm

[0086] By constructing a model predictive control framework and combining LSTM predicted trajectory and DQN initial speed sequence, distributed computing is used to optimize the speed sequence. MPC dynamically optimizes the speed based on real-time environmental conditions and predicted trajectory to adapt to short-term environmental changes. Distributed computing decomposes the optimization problem into sub-problems to meet the real-time control requirements of green buildings. This method can effectively improve control accuracy, thereby improving the effectiveness of adaptive speed control for fans in green buildings.

[0087] The optimized rotational speed sequence is input into the dynamic thermal balance model to generate the closed-loop control prediction state. By comparing the deviation between the closed-loop control prediction state and the actual environmental state, the LSTM network parameters and DQN strategy are dynamically adjusted, and the isolated forest algorithm is used to detect anomalies.

[0088] Furthermore, the process of dynamically adjusting the LSTM network parameters and DQN strategy by comparing the deviation between the predicted state of the closed-loop control and the actual environmental state, and using the isolated forest algorithm to detect anomalies includes:

[0089] The optimized speed sequence is input into the dynamic thermal balance model to generate the closed-loop control prediction state in time series format.

[0090] Compare the predicted state of the closed-loop control with the actual environmental state, calculate the weighted Euclidean distance as the deviation, and determine whether to trigger a model update.

[0091] Based on the deviation results, the LSTM network parameters and the reward function weights of the DQN strategy are dynamically adjusted in the cloud, the parameters are updated, and the data is sent to the edge devices.

[0092] The isolated forest algorithm is used to analyze the data collected by the IoT sensor network and detect abnormal states. If an abnormality is detected, a backup PID control strategy is triggered.

[0093] Furthermore, the optimized speed sequence is input into the dynamic heat balance model to update the ventilation heat flux term, which is related to the fan speed. The Runge-Kutta 4th order method is used to solve the problem on the edge device, with a prediction time domain of 5 minutes, to generate the closed-loop control prediction state.

[0094] Furthermore, the predicted state of the closed-loop control is compared with the corresponding temperature, humidity, and CO2 concentration in the actual environmental state; the deviation threshold is set to 0.8.

[0095] Furthermore, if the deviation exceeds the threshold, the cloud uses the PyTorch framework to fine-tune the parameters of the LSTM model online. The fine-tuning is performed every 5 minutes, and the updated parameters (weights and biases) are sent to the edge devices. At the same time, the cloud performs fine-tuning of the DQN reward function weights, updates the parameters, and sends them to the edge devices via MQTT.

[0096] Furthermore, temperature, CO2 concentration, fan vibration, and deviation were selected as the main analysis objects of the Isolation Forest algorithm, with 100 trees, an outlier ratio of 0.05, and a maximum sample size of 256. The average path length of each data point in the 100 trees was calculated to obtain the outlier score. If the outlier score exceeded 0.6 or the fluctuation of key features exceeded the limit, it was judged as an outlier.

[0097] By inputting the optimized speed sequence into the thermal balance model to generate the closed-loop predicted state, comparing the deviation to adjust the LSTM and DQN parameters, and using isolated forest to detect anomalies, the real-time control effect can be accurately evaluated, thereby assisting in the effectiveness of adaptive speed control of fans used in green buildings.

[0098] The method used in this embodiment is implemented through a cloud-edge collaborative architecture, where edge devices perform data processing and real-time control, while the cloud is responsible for model training and control optimization. This architecture enables cross-domain technology integration, dynamic adaptation and learning capabilities, as well as improved real-time performance and response speed, thereby enhancing the effectiveness of adaptive speed control for wind turbines used in green buildings.

[0099] This embodiment proposes an adaptive speed control method for wind turbines used in green buildings. First, indoor and outdoor environmental and operational data are collected and combined with building information model (BIM) data to generate a high-dimensional feature vector through multi-source heterogeneous data fusion. A dynamic thermal balance model is established, and an LSTM network is used to predict future environmental conditions, generating a control reference trajectory. An adaptive control strategy is constructed using a DQN network, with the high-dimensional feature vector and control reference trajectory as inputs, to generate an initial speed control sequence. A model predictive control framework is constructed, and the speed control sequence is optimized through distributed computing, combining the control reference trajectory and the initial speed control sequence, to generate an optimized speed sequence. By comparing the deviation between the predicted and actual closed-loop control states, the LSTM network parameters and the DQN strategy are adjusted, and anomalies are detected. This invention can improve the effectiveness of adaptive regulation in IoT control systems.

[0100] Example 2:

[0101] This invention proposes an adaptive speed control method for fans used in green buildings. To further verify the effectiveness of the proposed method, this embodiment applies the method to a green building (a 1000-square-meter office building equipped with a variable frequency centrifugal fan, rated power 500 watts, maximum speed 1800 rpm) to test the effectiveness of the method.

[0102] This invention selects historical data from a green building ventilation system over the past three years as the dataset for the methodology, using data from the first and second years as the training set and data from the third year as the validation set. The sampling rules are as follows: data from five days of each week are randomly selected on a weekly basis; each day, two sets of data are collected in the morning (8:00-10:00), noon (12:00-14:00), and evening (17:00-19:00), with each set lasting 30 minutes; the data covers the same type of fan (centrifugal fan, rated power 500 watts) to ensure consistency.

[0103] Four different control schemes were used to process the same historical test data to obtain the closed-loop control prediction state for each scheme. The historical test data included historical indoor and outdoor environmental data and historical fan operation status data. Building information model data could be directly obtained by calling the data. Then, a weighted Euclidean distance was calculated between the closed-loop control prediction state and the actual environmental state to obtain the deviation of each scheme. The deviation was then compared with the deviation threshold to obtain the proportion of each scheme's control result within the reasonable range.

[0104] The control schemes are as follows: the adaptive speed regulation control scheme proposed in this invention, namely "control reference trajectory generation + DQN adaptive control + model predictive control", is referred to as Scheme 1; the control reference trajectory generation process is removed, which is referred to as Scheme 2; the DQN adaptive control process is removed, which is referred to as Scheme 3; and the model predictive control process is removed, which is referred to as Scheme 4.

[0105] The results of the control scheme effectiveness test are shown in Table 3.

[0106] Table 3 Results of the effectiveness test of the control scheme

[0107]

[0108] The results in Table 3 show that the control scheme proposed in this invention performs better than other schemes in terms of effectiveness test results. This demonstrates that the "control reference trajectory generation + DQN adaptive control + model predictive control" scheme proposed in this invention can obtain more accurate control results, thereby improving the effectiveness of adaptive speed control for wind turbines used in green buildings.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive speed control method for fans used in green buildings, characterized in that, include: Indoor and outdoor environmental data and operational status data are collected through IoT sensor networks, and combined with building information model data. The Kalman filter algorithm is used to fuse multi-source heterogeneous data to generate high-dimensional feature vectors. A dynamic thermal balance model of the indoor environment is established, and the high-dimensional feature vector is processed by an LSTM network to predict the future environmental state and generate a control reference trajectory. An adaptive control strategy is constructed using a DQN network. The high-dimensional feature vector and the control reference trajectory are used as inputs, and the energy saving rate, PMV thermal comfort index and equipment life are used as multi-objective reward functions to generate an initial speed control sequence. A model predictive control framework is constructed, and by combining the control reference trajectory and the initial speed control sequence, the speed control sequence is optimized through distributed computing to generate an optimized speed sequence. The optimized rotational speed sequence is input into the dynamic thermal balance model to generate a closed-loop control prediction state. By comparing the deviation between the closed-loop control prediction state and the actual environmental state, the LSTM network parameters and DQN strategy are dynamically adjusted, and the isolated forest algorithm is used to detect anomalies.

2. The adaptive speed control method for a green building fan according to claim 1, characterized in that, The indoor and outdoor environmental data include: temperature, humidity, carbon dioxide concentration, and personnel density; the operating status data includes: rotational speed, power, vibration, and temperature; the building information model data includes geometric information, material property information, and functional information.

3. The adaptive speed control method for a green building fan according to claim 1, characterized in that, The process of generating the control reference trajectory includes: Based on the principles of building thermal dynamics, and combined with the room volume, material thermal conductivity, and air duct layout in the building information model data, a dynamic thermal balance model of the indoor environment is established to calculate the changes in real-time temperature, humidity, and CO2 concentration. The high-dimensional feature vector is input into an LSTM network to process the feature sequence within a preset time window, thereby obtaining the predicted temperature, humidity, and carbon dioxide concentration. A dynamic thermal balance model of the indoor environment is used to apply physical constraints to the LSTM prediction results, correct the prediction bias, and generate a control reference trajectory in time series format. The control reference trajectory is transmitted back from the cloud to the edge device via the MQTT protocol as input for adaptive control and model predictive control.

4. The adaptive speed control method for a green building fan according to claim 1, characterized in that, The process of constructing an adaptive control strategy using a DQN network and generating the initial speed control sequence includes: Construct a DQN network, define a discrete action space, and use the preprocessed high-dimensional feature vector and the control reference trajectory as inputs; Design a multi-objective reward function, define the optimization objective, calculate the reward value, and guide the DQN network to learn the optimal control strategy; A greedy strategy is used to train the DQN network in the cloud, and the experience replay mechanism is used to store state-action-reward data and update the neural network parameters. DQN inference is performed on the edge device to generate the initial speed control sequence based on the high-dimensional feature vector and the control reference trajectory input in real time.

5. The adaptive speed control method for a green building fan according to claim 1, characterized in that, The process of constructing a model predictive control framework, combining the control reference trajectory and the initial speed control sequence, and optimizing the speed control sequence through distributed computing to generate the optimized speed sequence includes: A model predictive control framework is constructed, which combines the control reference trajectory and the dynamic thermal balance model, and defines the prediction time domain, control time domain, and state transition function. Using the initial speed control sequence as the initial value, an optimization objective function is designed, taking into account the energy saving rate and PMV thermal comfort index, and constraints are set. The optimization problem is decomposed into sub-problems based on rooms or air ducts, and solved in parallel on edge devices through distributed computing to generate a local optimized speed sequence. The local optimization results are then aggregated in the cloud to generate an optimized speed sequence.

6. The adaptive speed control method for a green building fan according to claim 5, characterized in that, The state transition function uses the same formula as the indoor environment dynamic thermal balance model.

7. The adaptive speed control method for a green building fan according to claim 1, characterized in that, The process of dynamically adjusting the LSTM network parameters and DQN strategy by comparing the deviation between the predicted state of the closed-loop control and the actual environmental state, and using the isolated forest algorithm to detect anomalies includes: The optimized rotational speed sequence is input into the dynamic thermal balance model to generate the closed-loop control prediction state in time series format; Compare the predicted state of the closed-loop control with the actual environmental state, calculate the weighted Euclidean distance as the deviation, and determine whether to trigger a model update. Based on the deviation results, the LSTM network parameters and the reward function weights of the DQN strategy are dynamically adjusted in the cloud, the parameters are updated, and the data is sent to the edge devices. The isolated forest algorithm is used to analyze the data collected by the IoT sensor network and detect abnormal states. If an anomaly is detected, a backup PID control strategy is triggered.

Images