Multi-modal data fusion air conditioner optimization control method and system

Abstract

The invention discloses a multi-modal data fusion air conditioner optimization control method and system, and belongs to the technical field of intelligent building equipment control. According to the method, temperature, humidity, energy consumption, user behaviors and meteorological data are collected through a multi-source sensor, the data are subjected to dynamic window standardization processing and then input into a gated convolution LSTM network to predict the system state, an optimization strategy is generated in combination with an online reinforcement learning algorithm, multi-modal instructions are dynamically weighted and fused, and execution parameters are adjusted in real time through a feedback correction mechanism. The system comprises a multi-source sensing array, an edge computing unit, a strategy optimization engine and an intelligent execution controller. According to the method, through collaborative optimization of multi-modal spatial-temporal feature fusion and deep reinforcement learning, the problem of unbalance of energy efficiency and comfort is solved, the energy efficiency level and the user comfort of the air conditioning system are remarkably improved, and the method has the characteristics of real-time response and high stability, is suitable for intelligent air conditioning control of modern buildings and has wide application prospects.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent building equipment control, and particularly relates to a multi-modal data fusion air conditioner optimization control method and system. BACKGROUND

[0002] At present, the field of air conditioner intelligent control faces many technical bottlenecks, which seriously restricts the development of air conditioning systems in energy efficiency optimization, thermal comfort regulation and intelligent control.

[0003] From the perspective of balancing energy efficiency and comfort, the energy waste of air conditioning systems is more significant in building energy consumption. Traditional control methods cannot accurately match the user's demand for thermal comfort and the energy consumption of the system, resulting in a large amount of energy waste while meeting the user's comfort. For example, in the temperature regulation process, there are often excessive cooling or heating situations, which makes the indoor temperature deviate from the human comfort interval, not only causing invalid consumption of energy, but also causing many feedbacks from users about temperature fluctuations and humidity discomfort, and the user feedback rate is continuously high, which highlights the contradiction between energy efficiency optimization and thermal comfort regulation.

[0004] In terms of the coordinated use of multi-source sensing data, although various sensors are increasingly widely used in air conditioning systems, a large amount of information collected by sensors cannot be effectively integrated due to difficulties in matching space-time dimensions. The data collected by different types of sensors differ in time accuracy, spatial position and data format, making it difficult for these data to work together and provide strong support for the precise control of air conditioning systems, reducing the value of data utilization.

[0005] The real-time response capability of complex control algorithms is also insufficient. In modern building environments, air conditioning systems face sudden load changes such as personnel flow changes and outdoor weather condition mutations. However, existing complex control algorithms often have adjustment lags when facing these situations, and cannot timely and accurately adjust the operating state of the air conditioning system, making it difficult to meet the strict requirements of modern building environments for real-time and accuracy of air conditioning control.

[0006] Reviewing the existing technology, in the field of fuzzy control technology, Chinese patent CN201910557666.3 "Air conditioner control method based on fuzzy reasoning" constructs a two-dimensional fuzzy rule base and adjusts the compressor frequency only according to the temperature deviation and its rate of change. This single parameter input mode is too simple and does not fully consider other key factors such as humidity and energy consumption, resulting in a significant increase in energy consumption in actual application and the inability to achieve a balance between energy efficiency and comfort. Japanese patent JP2010256118A "Fuzzy control system for air conditioner" expands to a three-dimensional rule base and introduces humidity control dimension, but the complexity of the rules increases exponentially, making the reasoning process complex and causing reasoning delay problems, which reduces the response speed of the system.

[0007] In the aspect of predictive control technology, US patent US20170227345A1 "LSTM neural network energy consumption prediction system" uses a double-layer LSTM network to process historical temperature and humidity data to predict future load. However, its open-loop architecture causes prediction errors to accumulate and amplify, and over time, the deviation between the prediction result and the actual situation becomes larger and larger, seriously affecting the accuracy of control. European patent EP2988171A1 "Air conditioner load prediction method based on SVM" uses support vector machines to achieve cold load prediction, but the design of fixed sampling interval makes it unable to timely capture the rapid changes of load in transient scenarios, which performs poorly and cannot meet the accurate control needs of dynamic scenarios in actual application.

[0008] In the field of multi-sensor fusion technology, Japanese patent JP2020152666A "Multi-sensor linked air conditioner control device" sets a fixed weight fusion strategy for temperature, humidity, and CO2 concentration. However, in actual deployment, due to the influence of environmental factors, sensors will drift, and this static parameter allocation method cannot adapt to changes in sensor data, resulting in problems such as excessive ventilation, which affects the normal operation of the air conditioning system and energy utilization efficiency. Korean patent KR1020170088888 "Multi-parameter air conditioner control system" attempts a dynamic weight adjustment mechanism, but relies on manual intervention, resulting in response delay and inability to adjust control strategies in a timely manner according to environmental changes.

[0009] In the field of reinforcement learning technology, Chinese patent CN202110888888.8 "Multi-objective reinforcement learning method for air conditioning system" constructs a dual-objective optimization framework for energy consumption and comfort based on the PPO algorithm. However, this algorithm uses an offline update mechanism, which cannot adjust control strategies in a timely manner according to environmental changes when encountering sudden scenarios such as heavy rain, revealing the defect of control lag. US patent US20210089234 "Deep reinforcement learning air conditioner controller" uses the DDQN algorithm to achieve multi-parameter control, but the high deployment cost of GPU servers limits the commercial application of this technology, making it difficult to promote in the large-scale market.

[0010] In the aspect of edge computing technology, patent EP3512345B1 "Edge intelligent air conditioner control system" deploys Jetson platform to run LSTM network, improving real-time performance through high-frequency data sampling. However, the computing power of edge devices is limited, and when processing complex neural network inference, it is easy to cause inference delay, which cannot meet the real-time response requirements of the system. Patent CN202210123456 "Low-delay air conditioner control device" attempts to use a lightweight model compression scheme, which alleviates the pressure on edge computing to some extent, but results in accuracy loss, which reduces control stability and affects the control effect of the air conditioning system.

[0011] Innovation case WO2022006789A1 "Air conditioner digital twin system" realizes predictive maintenance through three-dimensional modeling. However, the complex calibration process of this technology not only increases the difficulty of technology implementation, but also leads to a significant increase in hardware costs, which seriously restricts its popularization and application in practice. Patent CN202310456789 "High-precision air conditioner simulation platform" couples CFD simulation to optimize air supply strategy, but the error problem has not been effectively solved in dynamic scene simulation, which cannot provide reliable basis for accurate control of the air conditioning system.

[0012] In summary, although the existing technology has made some progress in the field of air conditioning intelligent control, there are still defects in the balance between energy efficiency and comfort, collaborative use of multi-source sensor data, and real-time response of complex control algorithms, which cannot meet the urgent needs of modern building environments for efficient, comfortable, and intelligent control of air conditioning systems. SUMMARY

[0013] The present invention aims to solve the problems of energy efficiency and comfort imbalance, difficulty in collaborative use of multi-source sensor data, and insufficient real-time response capability of complex control algorithms in existing air conditioning intelligent control technology, and proposes a multi-modal data fusion air conditioning optimization control method and system.

[0014] TECHNICAL SOLUTION

[0015] The method of the present invention comprises the following steps:

[0016] 1. Data collection: Collect air conditioning system operation data through multi-source sensors, including temperature modal data, humidity modal data, energy consumption modal data, user behavior modal data, and environmental meteorological modal data.

[0017] 2. Data preprocessing: Normalize the collected multi-modal data, convert each modal data to a feature vector through standardization processing by subtracting the mean and dividing by the standard deviation. Adopt a sliding window mechanism to dynamically determine the window length, calculate the reference value based on the product of the maximum temperature difference threshold and the data volatility rate, and generate an integer window length value through an adjustment coefficient. The specific formula is:

[0018] [w=round(α×ΔR max ×V)], where (W) is the window length, (α) is the adjustment coefficient, and (ΔT) is the variable value. max (V) represents the maximum temperature difference threshold, and (V) represents the data volatility.

[0019] 3. State Prediction: A multi-step prediction deep reinforcement learning framework is constructed, employing a gated convolutional long short-term memory network for predicting the future state of the system in multiple steps. This network utilizes historical data features within a time sliding window for modeling and includes an input gate with convolutional kernel operations, a forget gate controlled by the sigmoid function, an output gate activated by tanh, and memory units with a gated update mechanism.

[0020] 4. Strategy Generation: An optimized control strategy is generated through an online adaptive reinforcement learning algorithm. This algorithm constructs an objective function based on the principle of maximizing discounted cumulative rewards. The simulated instructions include a comprehensive evaluation factor that combines energy efficiency and comfort indicators. The evaluation factor includes a real-time energy saving ratio based on baseline energy consumption, an exponential temperature deviation penalty term, and a quantified value for user preference matching, which are linearly combined using configurable weight coefficients. The specific formula is: [E=w1E1+w2E2+w3E3] where (E) is the comprehensive evaluation factor, (E1) is the real-time energy saving ratio based on baseline energy consumption, (E2) is the exponential temperature deviation penalty term, (E3) is the quantified value for user preference matching, and (w1,w2,w3) are weight coefficients.

[0021] 5. Command Fusion: A dynamic weighted fusion method is used to integrate multimodal prediction results. Real-time weight coefficients are assigned based on the historical accuracy of each modal predictor to generate a normalized optimal control command set. The weight coefficient adjustment adopts a softened version of the maximum confidence allocation principle, and dynamic weight allocation is completed through exponential calculation and normalization of the historical accuracy of each modal predictor. The specific formula is:

[0022] Among them, (w i ) represents the weight coefficients of the (i)th modality predictor, (Accuracy) i ) represents the historical accuracy of the (i)th modality predictor, (α) is the coefficient of the exponential operation, and (n) is the total number of modality predictors.

[0023] 6. Execution Control: A feedback correction mechanism with online adjustment capabilities is used to send control commands to the air conditioning actuators. This mechanism continuously optimizes parameters based on the gradient direction of the objective function. The objective function is a linear combination of energy efficiency ratio optimization terms, operating cost inverse terms, and parameter regularization terms. All sub-parameters have undergone range normalization. The specific formula is:

[0024] [J=weff J eff +w cost J cost +w reg J reg ], where (J) is the objective function, (J eff ) represents the energy efficiency ratio optimization term, (J cost ) is the inverse term of operating costs, (J reg ) is the parameter regularization term, (w eff ,w cost ,w reg ) represents the weighting coefficient of each item.

[0025] The system of the present invention includes:

[0026] Multi-source heterogeneous data acquisition device: equipped with temperature sensor, humidity detector, power analyzer, human body sensing unit and meteorological data interface, adopts time-division multiplexing communication protocol, and the sampling frequency meets the Nyquist criterion for joint sampling constraints of the three core parameters of temperature, humidity and power.

[0027] Data standardization processor: This module includes an embedded dynamic window parameter calculator and outlier filtering module to standardize and filter the collected data. The specific formula is as follows: Where (x) represents the original data, (μ) represents the mean, (σ) represents the standard deviation, and (x') represents the normalized data.

[0028] Multi-step state predictor: A gated convolutional long short-term memory network accelerator with parallel processing, responsible for predicting the future multi-step system state.

[0029] Policy optimization engine: Integrates reinforcement learning algorithm framework with online policy fusion unit, responsible for generating optimized control policies.

[0030] Intelligent actuator controller: Equipped with an adaptive adjustment feedback link, it is responsible for sending control commands to the air conditioning actuator and making online adjustments.

[0031] Beneficial effects

[0032] 1. Energy Efficiency Improvement: Through multimodal data fusion and collaborative optimization using deep reinforcement learning, this invention can significantly improve the energy efficiency of air conditioning systems. Compared to traditional control methods, this invention effectively reduces energy consumption and waste while meeting user comfort requirements.

[0033] 2. Enhanced Comfort: This invention, through dynamic weighted fusion of multimodal commands, can accurately match users' thermal comfort needs. By adjusting the operating parameters of the air conditioning system in real time, it ensures that indoor temperature and humidity remain within the human comfort range, thereby improving user comfort.

[0034] 3. Real-time response: The system of the present application has the characteristics of real-time response and strong stability. Through the feedback correction mechanism, the control instruction can be adjusted in time to cope with the sudden load changes such as personnel flow changes and outdoor weather condition mutations, ensuring that the running state of the air conditioning system is always in the optimal state.

[0035] 4. Improved data utilization efficiency: The present application solves the problem of data dimension fragmentation in the prior art through the collaborative use of multi-source sensing data. Through dynamic window standardization processing and gated convolution LSTM network, the data collected by different types of sensors can be effectively integrated to improve the utilization value of the data.

[0036] 5. Enhanced system stability: The system of the present application can continuously optimize the control strategy and parameters through online adaptive reinforcement learning algorithm and feedback correction mechanism, ensuring the long-term running stability of the system. Under different environments and user demands, the system can maintain good running state, providing reliable guarantee for the intelligent control of air conditioning system. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings of the present application will be briefly described below. Those skilled in the art can obtain other embodiments based on these drawings, and these drawings do not exceed the scope of the present application.

[0038] Figure 1 : The flow chart of the multi-modal data fusion air conditioning optimization control method of the embodiment of the present application;

[0039] Figure 2 : The system architecture diagram of the multi-modal data fusion air conditioning optimization control system of the embodiment of the present application;

[0040] Figure 3 : The data fusion process diagram of the embodiment of the present application;

[0041] Figure 4 : The control instruction generation process diagram of the embodiment of the present application;

[0042] Figure 5 : The air conditioning optimization control system application scenario diagram of the embodiment of the present application;

[0043] Figure 6 : The data preprocessing process diagram of the embodiment of the present application;

[0044] Figure 7 : The weight adjustment diagram of the embodiment of the present application;

[0045] Figure 8 : The fuzzy control algorithm diagram of the embodiment of the present application;

[0046] Figure 9 : User behavior prediction schematic diagram for an embodiment of the present application;

[0047] Figure 10 : Energy consumption and comfort optimization effect schematic diagram for an embodiment of the present application;

[0048] Figure 11 : Gated convolution long short-term memory network structure schematic diagram for an embodiment of the present application;

[0049] Figure 12 : Energy consumption and comfort optimization effect schematic diagram for an embodiment of the present application;

[0050] Figure 13 : Energy consumption and comfort optimization effect schematic diagram for an embodiment of the present application. DETAILED DESCRIPTION

[0051] The present application aims to solve the problems of energy efficiency and comfort imbalance, difficulty in collaborative use of multi-source sensing data, and insufficient real-time response capability of complex control algorithms in existing air conditioner intelligent control technology, and proposes a multi-modal data fusion air conditioner optimization control method and system.

[0052] The present embodiment provides a multi-modal data fusion air conditioner optimization control method, the specific steps are as follows:

[0053] 1. Data acquisition (S01): Collect air conditioner system operation data through multi-source sensors, including temperature modal data, humidity modal data, energy consumption modal data, user behavior modal data and environmental meteorological modal data, as shown in Figure 3 . In specific implementation, a variety of sensors can be used, such as temperature sensors, humidity sensors, power analyzers, human body sensors and meteorological data interfaces, to obtain comprehensive system operation data. In the figure, Figure 3 , 111 is the temperature modal data, 112 is the humidity modal data, 113 is the energy consumption modal data, 114 is the user behavior modal data, 115 is the environmental meteorological modal data, and 116 is the fusion unit.

[0054] Working process: Temperature sensors, humidity sensors, power analyzers, human body sensors and meteorological data interfaces collect temperature, humidity, energy consumption, user behavior and environmental meteorological data in the operation of the air conditioner system. These data are integrated by the fusion unit to form a multi-modal data set, providing a basis for subsequent data processing and analysis.

[0055] 2. Data preprocessing (S02): Normalize the collected multi-modal data, convert each modal data to a feature vector through standardization processing of subtracting the mean and dividing by the standard deviation, as shown in Figure 6The length of the window is dynamically determined using a sliding window mechanism. The reference value is calculated based on the product of the maximum temperature difference threshold and the data volatility rate. An adjustment coefficient is used to generate an integer window length value. The formula is as follows:

[0056] L = a x AT max x s where L is the window length, a is the adjustment coefficient, AT is the maximum temperature difference threshold, and s is the data volatility rate. max

[0057] Reference signs: Figure 6 In the figure, 61 is the original data, 62 is the dynamic window processing, 63 is the standardization, and 64 is the normalization.

[0058] Working process: The original data is processed by the dynamic window, and the data of different time periods is divided into multiple windows. Then the data in each window is standardized, which means subtracting the mean and dividing by the standard deviation, so that the distribution characteristics are normalized. For example, if the collected temperature data is 30°C, the mean is 25°C, and the standard deviation is 2°C, the standardized value is (30-25) / 2=2.5. After normalization, the value range of the data feature vector is unified to a standard interval, which is convenient for subsequent model training and prediction.

[0059] 3. State prediction (S03): A multi-step prediction deep reinforcement learning framework is constructed, and a gated convolutional long short-term memory network is used for future multi-step system state prediction, as shown in Figure 11 The network uses historical data features in a time sliding window for modeling, including an input gate with convolution kernel operation, a forget gate controlled by a sigmoid function, an output gate with tanh activation, and a memory unit with a gated update mechanism.

[0060] Reference signs: Figure 11 In the figure, 111 is the input gate, 112 is the forget gate, 113 is the output gate, and 114 is the memory unit.

[0061] Working process: The input gate extracts the local spatial features of the input data through convolution kernel operation and decides which information needs to be written to the memory unit. The forget gate uses a sigmoid function to control the retention ratio of historical information in the memory unit and decides which information needs to be forgotten. The output gate uses tanh activation to generate the hidden state at the current time, which is part of the prediction result. The memory unit realizes information flow control through a gated update mechanism and point multiplication operation, integrates the information of the input gate, forget gate, and output gate, and thus predicts the future system state. For example, when predicting the future temperature of an air conditioning system, the model analyzes the historical temperature data, humidity, energy consumption, and other factors, and gives a prediction of the future temperature after comprehensive consideration.

[0062] ​4. Strategy Generation (S04): An optimized control strategy is generated through an online adaptive reinforcement learning algorithm. This algorithm constructs the objective function based on the principle of maximizing discounted cumulative rewards. The simulated commands include comprehensive evaluation factors of energy efficiency and comfort indicators, such as... Figure 12 As shown. In specific implementation, the evaluation factors include the real-time energy saving ratio based on benchmark energy consumption, an exponential temperature deviation penalty term, and a user preference matching metric, which are linearly combined through configurable weighting coefficients, as shown in the following formula: R = β1 × E s +β2×T d +β3×U m Where R is the comprehensive evaluation factor, E s T is the real-time energy saving ratio based on baseline energy consumption. d U is an exponential temperature deviation penalty term. m Metrics are matched to user preferences, where β1, β2, and β3 are weighting coefficients.

[0063] Figure label: Figure 12 In the diagram, 121 represents the reinforcement learning algorithm framework, 122 represents the online policy fusion unit, and 123 represents the objective function.

[0064] Working Process: The reinforcement learning algorithm framework continuously optimizes the control strategy based on the current system state and existing experience data. The online policy fusion unit merges policies from different sources to generate the final control command. The objective function, through the principle of maximizing discounted cumulative rewards, comprehensively considers factors such as energy efficiency indicators, comfort indicators, and user preferences, guiding the algorithm to generate a control strategy that is both energy-efficient and comfortable. For example, when the air conditioning system is running, the algorithm dynamically adjusts parameters such as compressor frequency and fan speed based on real-time data such as the indoor-outdoor temperature difference and occupant activity to achieve optimal energy saving and comfort.

[0065] 5. Command Fusion (S05): A dynamic weighted fusion method is used to integrate multimodal prediction results. Real-time weight coefficients are assigned based on the historical accuracy of each modal predictor to generate a normalized optimal control command set, such as... Figure 7 As shown. In practice, the weight coefficient adjustment adopts a softened version of the maximum confidence allocation principle. Dynamic weight allocation is achieved through exponential calculation and normalization of the historical accuracy of each modality predictor, as shown in the following formula: Among them, w i Acc represents the weight coefficients of the i-th modality predictor. i Let η be the historical accuracy of the i-th modal predictor, η be the coefficient of the exponential operation, and N be the total number of modal predictors.

[0066] Figure label: Figure 7 In the table, 71 is for historical accuracy calculation, 72 is for exponential operation, and 73 is for weight allocation.

[0067] Working process: The system first calculates the historical accuracy of each modal predictor, and then converts the accuracy into weight coefficients through exponential operation and normalization processing. The higher the weight coefficient, the more reliable the prediction result of the predictor, and the greater the influence in decision-making. For example, if the historical accuracy of the temperature predictor is higher than that of the humidity predictor, the temperature prediction result will be given greater weight when generating the control instruction.

[0068] 6. Perform control (S06): The control instruction is issued to the air conditioner actuator through a feedback correction mechanism with online adjustment capability, which continuously optimizes the parameters based on the gradient direction of the objective function, as shown in the formula: Figure 13 When implemented, the objective function includes a linear combination of the energy efficiency ratio optimization term, the operating cost inverse term, and the parameter regularization term, and each sub-item parameter is subjected to value range normalization processing, as follows: Where J is the objective function, E save is the energy efficiency ratio optimization term, E total is the total energy consumption, C is the operating cost inverse term, Reg is the parameter regularization term, and γ1, γ2, γ3 are the weight coefficients of each sub-item.

[0069] Reference signs: Figure 13 Where 131 is the adaptive adjustment feedback link, 132 is the proportional gain adjustment, 133 is the integral time adjustment, and 134 is the actuator control signal.

[0070] Working process: The actuator adjusts the operating parameters of the air conditioning system according to the control instruction, such as compressor frequency, fan speed, valve opening, etc. The feedback correction mechanism monitors the system operating state in real time, compares the actual operating data with the target value, calculates the deviation and adjusts the control parameters. For example, if the room temperature is lower than the set value, the feedback correction mechanism will adjust the compressor frequency to reduce it, thereby reducing the cooling output, so that the temperature gradually rises to the set value.

[0071] The embodiment provides an air conditioning system optimization control system based on multi-modal data fusion, and the specific structure is as follows:

[0072] 1. Multi-source heterogeneous data acquisition device (1): temperature sensor, humidity detector, power analyzer, human body sensing unit and weather data interface are configured, time division multiplexing communication protocol is adopted, sampling frequency meets the joint sampling constraint of temperature, humidity and power three types of core parameters according to Nyquist criterion, as shown in Figure 2 .

[0073] Reference signs: Figure 2 Where 1 is the multi-source heterogeneous data acquisition device, 2 is the data standardization processor, 3 is the multi-step state predictor, 4 is the strategy optimization engine, and 5 is the intelligent execution controller.

[0074] Working process: The multi-source heterogeneous data acquisition device collects temperature, humidity, power and other data through various sensors and interfaces, and transmits the data to the data standardization processor through time division multiplexing communication protocol. The sampling frequency is strictly controlled to ensure that the collected data can accurately reflect the real-time running state of the air conditioning system and meet the requirements of the Nyquist criterion.

[0075] 2. Data standardization processor (2): embedded dynamic window parameter calculator and outlier filtering module, standardizing and filtering the collected data, as shown in Figure 6 . In specific implementation, the normalization preprocessing method in the above embodiment 1 can be used.

[0076] Reference signs: Figure 6 In the figure, 61 is the original data, 62 is the dynamic window processing, 63 is the standardization, and 64 is the normalization.

[0077] Working process: The dynamic window parameter calculator divides the original data into different windows, and the data in each window is processed by the outlier filtering module to remove abnormal data points. Then, the standardization processing is performed to convert the data into feature vectors, providing accurate data input for subsequent state prediction and strategy generation.

[0078] 3. Multi-step state predictor (3): contains a parallel processing gated convolution long short-term memory network accelerator, responsible for predicting future multi-step system state, as shown in Figure 11 .

[0079] Reference signs: Figure 11 In the figure, 111 is the input gate, 112 is the forget gate, 113 is the output gate, and 114 is the memory cell.

[0080] Working process: The gated convolution long short-term memory network accelerator uses parallel processing technology to quickly model the historical data features in the time sliding window. Through the cooperative operation of the input gate, the forget gate, the output gate and the memory cell, the future multi-step system state is accurately predicted.

[0081] 4. Strategy optimization engine (4): integrates reinforcement learning algorithm framework and online strategy fusion unit, responsible for generating optimized control strategy, as shown in Figure 12 . In specific implementation, the strategy generation method in the above embodiment 1 can be used.

[0082] Reference signs: Figure 12 In the figure, 121 is the reinforcement learning algorithm framework, 122 is the online strategy fusion unit, and 123 is the objective function.

[0083] Working process: Reinforcement learning algorithm framework and online policy fusion unit work together to generate optimized control strategies. The algorithm framework continuously optimizes policy parameters based on real-time system state and historical data. The online policy fusion unit integrates different policy suggestions to generate the final control instructions, ensuring that the air conditioning system operates both energy-efficiently and comfortably.

[0084] 5. Intelligent execution controller (5): Equipped with adaptive adjustment feedback link, responsible for issuing control instructions to air conditioning execution mechanism and performing online adjustment, as shown in Figure 13 . In specific implementation, the execution control method in the above embodiment 1 can be used.

[0085] Reference signs: Figure 13 , 131 is the adaptive adjustment feedback link, 132 is the proportional gain adjustment, 133 is the integral time adjustment, and 134 is the execution mechanism control signal.

[0086] Working process: The intelligent execution controller accurately sends control instructions to the air conditioning execution mechanism through the adaptive adjustment feedback link. At the same time, according to the changes of system running state, automatically adjusts the control parameters such as proportional gain and integral time, to adapt to different environments and user demands, ensuring the stable operation of the air conditioning system.

[0087] This embodiment further expands the application scenarios of the above-mentioned multi-modal data fusion air conditioning optimization control method and system, as follows:

[0088] 1. Data fusion process Figure 3 ): Show the collection and fusion process of multi-source data, including temperature modal data, humidity modal data, energy consumption modal data, user behavior modal data and environmental meteorological modal data input and fusion result output.

[0089] Reference signs: Figure 3 , 111 is the temperature modal data, 112 is the humidity modal data, 113 is the energy consumption modal data, 114 is the user behavior modal data, 115 is the environmental meteorological modal data, and 116 is the fusion unit.

[0090] Working process: After various modal data is collected, it is integrated through the fusion unit. The fusion process fully considers the spatio-temporal correlation and complementarity between different data, so that the fused data can fully reflect the running state of the air conditioning system and user demand.

[0091] 2. Control instruction generation process Figure 4 ): Show the generation of optimized control strategies and the fusion process of multi-modal prediction results, including the interaction of strategy optimization engine, instruction fusion module and air conditioning execution mechanism.

[0092] Reference signs:Figure 4 In this case, 41 is the policy optimization engine, 42 is the instruction fusion module, and 43 is the air conditioning actuator.

[0093] Working process: The policy optimization engine generates multiple control strategy suggestions based on the multi-modal prediction results. The instruction fusion module integrates these strategy suggestions to generate the final control instructions. The air conditioning actuator receives the control instructions and adjusts the system operating parameters to achieve precise control of the air conditioning system.

[0094] 3. Application scenarios of air conditioning optimization control system Figure 5 ): Show the application scenarios of air conditioning system in indoor and outdoor environment, including indoor environment, outdoor environment and deployment and running environment of air conditioning system.

[0095] Reference signs: Figure 5 In this case, 51 is the indoor environment, 52 is the outdoor environment, and 53 is the air conditioning system.

[0096] Working process: In the indoor environment, the air conditioning system adjusts according to the user's comfort demand and indoor environment parameters (such as temperature, humidity, etc.). In the outdoor environment, the air conditioning system considers outdoor meteorological data (such as temperature, humidity, wind speed, etc.), as well as the interaction with the indoor environment, such as indoor and outdoor temperature difference, heat exchange, etc., to adjust the operation strategy.

[0097] 4. Data preprocessing process Figure 6 ): Show the whole picture of data preprocessing, from raw data to normalized data processing flow, including dynamic window processing and standardization of specific methods.

[0098] Reference signs: Figure 6 In this case, 61 is the raw data, 62 is the dynamic window processing, 63 is the standardization, and 64 is the normalization.

[0099] Working process: Raw data is processed by dynamic window, which divides data into multiple windows. Then the data in each window is standardized to eliminate the influence of data dimension, so that data of different modalities can be analyzed and processed on the same scale. Figure 7 5. Weight adjustment

[0100] ): Show the whole picture of weight adjustment, from historical accuracy to weight distribution process, including exponential operation and normalization formula.

[0101] Reference signs: Figure 7 In this case, 71 is the historical accuracy calculation, 72 is the exponential operation, and 73 is the weight distribution.

[0102] Working process: The system dynamically adjusts the weight coefficients based on the historical accuracy of each modal predictor through exponential operation and normalization processing. The adjustment of the weight coefficients enables the system to flexibly allocate decision weights according to the reliability of the predictors, improving the accuracy of the decision.

[0103] 6. Fuzzy control algorithm Figure 8 ): Show the whole picture of fuzzy control algorithm, from input parameters to output control quantity, including fuzzification, rule base and inference engine specific operation.

[0104] Reference signs: Figure 8 In the figure, 81 is the input parameter, 82 is the fuzzification, 83 is the rule base, 84 is the inference engine, 85 is the defuzzification, and 86 is the output control quantity.

[0105] Working process: The input parameters are converted into fuzzy sets after fuzzification. The rule base stores a variety of fuzzy rules, and the inference engine infers the fuzzy control quantity based on the input fuzzy set and fuzzy rules. Then through the defuzzification process, the fuzzy control quantity is converted into the actual control signal.

[0106] 7. User behavior prediction Figure 9 ): Show the whole picture of user behavior prediction, from historical data to future behavior prediction process, including specific algorithm of prediction model.

[0107] Reference signs: Figure 9 In the figure, 91 is the historical behavior data, 92 is the prediction model, and 93 is the future behavior pattern.

[0108] Working process: The prediction model establishes a prediction model of user behavior pattern based on the user's historical behavior data through machine learning algorithm. Then according to the current environmental parameters and user behavior data, it predicts the possible behavior demand of future users.

[0109] 8. Energy consumption and comfort optimization effect Figure 10 ): Show the comparison results of optimization control mode and traditional control mode in energy consumption and comfort, including specific data of energy consumption comparison and comfort comparison.

[0110] Reference signs: Figure 10 In the figure, 101 is the traditional control mode, 102 is the optimization control mode, 103 is the energy consumption comparison, and 104 is the comfort comparison.

[0111] Working process: Through comparative experiments, under the same environmental and user demand conditions, the traditional control mode and the optimization control mode are used respectively to run the air conditioning system. Statistical analysis of energy consumption and user comfort data under two modes shows the significant advantages of optimization control mode in energy saving and comfort.

[0112] 9. Gated Convolutional Long Short-Term Memory Network Structure Figure 11 ) : Show the specific structure of the gated convolutional long short-term memory network, including the detailed design of the input gate, the forget gate, the output gate and the memory unit.

[0113] Reference signs: Figure 11 In the figure, 111 is the input gate, 112 is the forget gate, 113 is the output gate, and 114 is the memory unit.

[0114] Working process: The input gate controls the degree of input data entering the memory unit, the forget gate controls the degree of retention of old information in the memory unit, and the output gate controls the degree of output of information in the memory unit. The memory unit processes time series data efficiently through the gating mechanism and convolution operation, extracts long-term dependencies and spatial features.

[0115] 10. Strategy Optimization Engine Architecture Figure 12 ) : Show the detailed architecture of the strategy optimization engine, including the reinforcement learning algorithm framework, the online policy fusion unit and the construction of the objective function.

[0116] Reference signs: Figure 12 In the figure, 121 is the reinforcement learning algorithm framework, 122 is the online policy fusion unit, and 123 is the objective function.

[0117] Working process: The reinforcement learning algorithm framework optimizes the control strategy through continuous trial and error and learning. The online policy fusion unit fuses different strategies to generate the final control instruction. The objective function considers multiple factors such as energy efficiency and comfort to guide the system to develop in the optimal direction.

[0118] 11. Intelligent Execution Controller Control Logic Figure 13 ) : Show the control logic of the intelligent execution controller, including the specific mechanism of adaptive adjustment feedback link, proportional gain adjustment and integral time adjustment.

[0119] Reference signs: Figure 13 In the figure, 131 is the adaptive adjustment feedback link, 132 is the proportional gain adjustment, 133 is the integral time adjustment, and 134 is the actuator control signal.

[0120] Working process: The adaptive adjustment feedback link automatically adjusts the control parameters according to the changes in the system running state. The proportional gain adjustment adjusts the control strength in real time according to the size of the error. The integral time adjustment further optimizes the control strategy according to the accumulation of errors, so that the air conditioning system always maintains the best running state.

[0121] Through the above-mentioned embodiments, the air conditioning system can significantly improve the energy efficiency level and user comfort, has the characteristics of real-time response and stability, is suitable for intelligent air conditioning control of modern buildings, and has a wide application prospect.

[0122] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0123] The above-mentioned embodiments are only used to illustrate the technical solutions of the present application, but not limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A multimodal data fusion air conditioning optimization control method, characterized in that, Includes the following steps: S01: Collect air conditioning system operation data through multi-source sensors, including temperature modal data, humidity modal data, energy consumption modal data, user behavior modal data, and environmental meteorological modal data; S02: Perform normalization preprocessing on the multimodal data, converting each modal data into a feature vector by subtracting the mean and dividing by the standard deviation; S03: Construct a multi-step prediction deep reinforcement learning framework, and use a gated convolutional long short-term memory network to predict the future multi-step system state. The network uses historical data features within a time sliding window for modeling. S04: An optimized control strategy is generated through an online adaptive reinforcement learning algorithm. The algorithm constructs an objective function based on the principle of maximizing discounted cumulative rewards. The simulated commands include a comprehensive evaluation factor that combines energy efficiency and comfort indicators. S05: The dynamic weighted fusion method is used to integrate the multimodal prediction results, and real-time weight coefficients are allocated according to the historical accuracy of each modal predictor to generate a normalized optimal control instruction set; S06: Control commands are sent to the air conditioning actuator through a feedback correction mechanism with online adjustment capability. This mechanism continuously optimizes parameters based on the gradient direction of the objective function.

2. The method according to claim 1, characterized in that, The normalization preprocessing described in step S02 adopts a sliding window mechanism. The dynamic determination of the window length includes: calculating the baseline value based on the product of the maximum temperature difference threshold and the data volatility, and generating an integer window length value through an adjustment coefficient.

3. The method according to claim 1, characterized in that, The gated convolutional long short-term memory network described in step S03 includes: 3a: Input gates with convolutional kernel operations are used to capture local spatial features; 3b: The forget gate, controlled by the sigmoid function, determines the proportion of state information retained; 3c: Use an output gate activated by tanh to generate the hidden state at the current moment; 3d: A memory unit with a gated update mechanism that controls information flow through dot product operations.

4. The method according to claim 1, characterized in that, The evaluation factors mentioned in step S04 include: 4a: Real-time energy saving ratio based on baseline energy consumption; 4b: Exponential temperature deviation penalty; 4c: User preference matching metrics; The three types of evaluation indicators are linearly combined using configurable weight coefficients.

5. The method according to claim 1, characterized in that, The weight coefficient adjustment in step S05 adopts a softened version of the maximum confidence allocation principle, and the dynamic weight division is completed by exponential calculation and normalization of the historical accuracy of each modality predictor.

6. The method according to claim 1, characterized in that, The objective function described in step S06 includes a linear combination of an energy efficiency ratio optimization term, an inverse operating cost term, and a parameter regularization term, and all sub-parameters have undergone value range normalization.

7. An optimization control system for an air conditioning system based on multimodal data fusion, characterized in that, include: A multi-source heterogeneous data acquisition device, equipped with a temperature sensor, humidity detector, power analyzer, human body sensing unit and meteorological data interface; 7a: Data standardization processor with embedded dynamic window parameter calculator and outlier filtering module; 7b: Multi-step state predictor with a gated convolutional long short-term memory network accelerator for parallel processing; 7c: Policy optimization engine, integrating reinforcement learning algorithm framework and online policy fusion unit; 7d: Intelligent execution controller, equipped with an adaptive adjustment feedback link; 7e: The system implements the control logic of the method as described in any one of claims 1 to 6.

8. The system according to claim 7, characterized in that, The multi-source heterogeneous data acquisition device adopts a time-division multiplexing communication protocol, and its sampling frequency meets the Nyquist criterion for the joint sampling constraints of the three core parameters of temperature, humidity, and power.

9. The system according to claim 7, characterized in that, The policy optimization engine is built as a heterogeneous computing architecture, in which the main control processor is responsible for policy evaluation and weight calculation, and the coprocessor is dedicated to running the deep neural network forward propagation and parameter update program.

10. The system according to claim 7, characterized in that, The intelligent execution controller is equipped with a composite control algorithm module, whose parameter tuning mechanism includes proportional gain error response mode adjustment and integral time cumulative error adaptive adjustment function.

Images