Meteorological big data-based air water production intelligent prediction and adjustment method and system

Abstract

The invention provides an air water production intelligent prediction and adjustment method and system based on meteorological big data, and relates to the technical field of air water production, and the method comprises the steps: collecting data through a meteorological sensor, constructing an enhanced feature space, fusing multi-domain knowledge through a migration cross-domain learning network, and predicting the water production through combining with a deep neural network. A control strategy is generated based on a deep reinforcement learning model, optimal control parameters are selected through multi-target Bayesian optimization and a dynamic decision algorithm, the air water production amount can be accurately predicted, intelligent adjustment is achieved, the water production efficiency is improved, energy consumption is reduced, and high environmental adaptability and robustness are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of air-to-water production, and in particular to a method and system for intelligently predicting and regulating air-to-water production based on meteorological big data. Background Art

[0002] Air-to-water technology is an innovative technology for extracting water from the air. It converts water vapor in the air into liquid water through the principle of condensation and cooling. It has important application value in water supply in water-scarce areas, emergency rescue and other fields. The efficiency of air-to-water is mainly affected by meteorological factors such as ambient temperature and relative humidity. Reasonable prediction and adjustment of water production parameters are of great significance to improving water production efficiency.

[0003] At present, air-to-water systems usually adopt fixed parameter control schemes based on empirical rules, or use simple feedback control strategies for parameter adjustment. With the development of meteorological monitoring technology and artificial intelligence technology, attempts are made to apply machine learning methods to water production parameter optimization to improve system efficiency.

[0004] However, the existing air-to-water control methods still rely on a single meteorological data source for prediction and fail to fully utilize the complementarity of multi-source meteorological data, resulting in limited prediction accuracy and poor adaptability to abnormal conditions such as extreme weather.

[0005] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for intelligent prediction and regulation of air-to-water production based on meteorological big data, which can at least solve some of the problems existing in the prior art.

[0007] A first aspect of an embodiment of the present invention provides an intelligent prediction and regulation method for air-to-water production based on meteorological big data, comprising:

[0008] Meteorological monitoring data is collected through meteorological sensors and standardized to generate a standardized meteorological feature matrix. Periodic feature vectors and trend feature vectors are extracted through time series decomposition and an enhanced feature space is constructed.

[0009] Construct a transfer cross-domain learning network and add the pre-trained model knowledge in the fields of refrigeration and air conditioning and weather forecasting to the air-to-water prediction task through the knowledge transfer algorithm. Combine knowledge distillation to integrate expert knowledge from multiple source domains into the deep neural network.

[0010] Based on the enhanced feature space, a deep neural network is trained in combination with a back propagation algorithm to output the predicted water production volume hour by hour in the next day, the predicted water production volume is used as the input of the state space and the action space is configured, a reward function is constructed according to the water production efficiency and energy consumption ratio, and a deep reinforcement learning model is constructed based on the state space, the action space and the reward function;

[0011] Dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework and construct training samples corresponding to extreme meteorological conditions through an adversarial sample generator, optimize the deep reinforcement learning model based on the training samples to obtain a set of candidate water production parameter control strategies, execute the water production process based on the candidate water production parameter control strategy set and set the optimization target;

[0012] A multi-objective Bayesian optimizer is used for joint optimization to generate a Pareto optimal solution set, and the optimal control parameters are selected through a dynamic decision-making algorithm based on a confidence boundary. The update step size of the deep reinforcement learning model is dynamically adjusted based on the optimal control parameters to obtain the optimal water production strategy.

[0013] In an optional embodiment,

[0014] Meteorological monitoring data is collected through meteorological sensors and standardized to generate a standardized meteorological feature matrix. Periodic feature vectors and trend feature vectors are extracted through time series decomposition and an enhanced feature space is constructed, including:

[0015] Collect meteorological monitoring data through meteorological sensors, perform standardization on the meteorological monitoring data, and arrange the standardized meteorological monitoring data in time series to generate a standardized meteorological feature matrix;

[0016] Performing time series decomposition on the standardized meteorological feature matrix, constructing a trajectory matrix and performing singular value decomposition, and decomposing the standardized meteorological feature matrix into a periodic component and a trend component, extracting a periodic feature vector and a trend feature vector of the standardized meteorological feature matrix;

[0017] An enhanced feature space is constructed based on the periodic feature vector and the trend feature vector, wherein the enhanced feature space is used to characterize the periodic variation law and long-term variation trend of the meteorological monitoring data.

[0018] In an optional embodiment,

[0019] Construct a migration cross-domain learning network and add the pre-trained model knowledge in the refrigeration and air conditioning field and the weather forecast field to the air-to-water prediction task through the knowledge transfer algorithm. Combine knowledge distillation to integrate expert knowledge from multiple source domains into the deep neural network, including:

[0020] Constructing a transfer cross-domain learning network, the transfer cross-domain learning network includes a feature extractor, a task-specific classifier and a domain discriminator, the feature extractor adopts a convolutional neural network structure with residual connection, the domain discriminator adopts a three-layer fully connected network structure, and the feature extractor is used to extract feature vectors from the refrigeration and air-conditioning field and the weather forecast field;

[0021] Acquire feature statistical information based on the feature vector and input it into a kernel function generation network to obtain a kernel bandwidth parameter and a kernel function combination weight, identify the source domain of the feature vector based on a domain adversarial loss function to obtain a domain adversarial loss value, construct a maximum average difference metric based on the kernel bandwidth parameter and the kernel function combination weight, and calculate the difference between the feature vectors in the refrigeration and air conditioning field and the weather forecast field to obtain a maximum average difference value;

[0022] For each feature vector, an attention mechanism is used to calculate the importance of the pre-trained model knowledge in the refrigeration and air-conditioning field and the weather forecast field to obtain a corresponding attention weight, and the attention weight is multiplied by the feature vector and superimposed to obtain a fusion feature;

[0023] Based on the knowledge distillation loss function, the expert knowledge in the fusion feature is transferred to the deep neural network to obtain a knowledge distillation loss value, and the knowledge distillation loss value, the domain adversarial loss value and the maximum average difference value are weightedly combined to obtain an overall loss function;

[0024] An alternating optimization strategy is adopted to train the transfer cross-domain learning network, the domain discriminator parameters corresponding to the domain discriminator are fixed and the feature extractor parameters corresponding to the feature extractor are optimized based on the overall loss function, the feature extractor parameters are fixed and the domain discriminator parameters are optimized based on the output of the domain discriminator, and the parameter optimization process is iteratively performed through the back propagation algorithm until the overall loss function converges to obtain a trained transfer cross-domain learning network.

[0025] In an optional embodiment,

[0026] Acquiring feature statistical information based on the feature vector and inputting it into a kernel function generation network to obtain a kernel bandwidth parameter and a kernel function combination weight, performing source domain identification on the feature vector based on a domain adversarial loss function to obtain a domain adversarial loss value, constructing a maximum average difference metric based on the kernel bandwidth parameter and the kernel function combination weight, and calculating the feature vector difference between the refrigeration and air conditioning field and the weather forecast field to obtain the maximum average difference value includes:

[0027] Based on the feature vector, a feature pyramid structure is constructed in multiple network layers of a feature extractor, wherein the feature pyramid structure includes a shallow feature vector, a middle feature vector and a deep feature vector, wherein the shallow feature vector includes local texture information, the middle feature vector includes local semantic information, and the deep feature vector includes global semantic information;

[0028] Calculating the mutual information between the shallow feature vector, the middle feature vector and the deep feature vector to obtain an inter-layer correlation score, and selecting the feature level combination with the strongest discriminability based on the inter-layer correlation score to obtain a combined feature vector;

[0029] Acquire the mean and variance of the combined feature vector to construct feature statistical information, input the feature statistical information into a kernel function generation network, and the kernel function generation network outputs a kernel bandwidth parameter and a kernel function combination weight;

[0030] Calculating the inter-class distance and intra-class distance of the combined feature vector in the refrigeration and air-conditioning field to obtain the feature distance in the refrigeration and air-conditioning field, and calculating the feature variance of the combined feature vector in the weather forecast field to obtain the feature variance in the weather forecast field;

[0031] Calculating the feature dimension importance based on the feature distance in the refrigeration and air conditioning field and the feature variance in the weather forecast field, and weighting the feature dimension importance with the combined feature vector to obtain a weighted feature vector;

[0032] The weighted feature vector is input into a domain discriminator, and the source domain of the weighted feature vector is identified based on a domain adversarial loss function to obtain a domain adversarial loss value. A maximum average difference metric is constructed based on the kernel bandwidth parameter and the kernel function combination weight. The maximum average difference metric is used to calculate the difference between the weighted feature vectors in the refrigeration and air-conditioning field and the meteorological forecast field to obtain a maximum average difference value.

[0033] In an optional embodiment,

[0034] Based on the enhanced feature space, the deep neural network is trained in combination with the back propagation algorithm and the predicted water production volume is output hour by hour in the next day. The predicted water production volume is used as the input of the state space and the action space is configured. The reward function is constructed according to the water production efficiency and energy consumption ratio. The deep reinforcement learning model is constructed based on the state space, action space and reward function, including:

[0035] A deep neural network is constructed based on the enhanced feature space, wherein the deep neural network includes an input layer, three hidden layers and an output layer, wherein the hidden layer uses a rectified linear unit activation function; a back propagation algorithm is used to train the deep neural network to obtain hourly predicted water production in the next day;

[0036] The predicted water production volume is constructed as the input of the state space, and an action space including compressor frequency, fan speed and system start / stop state is configured; a reward function is constructed based on water production efficiency and energy consumption ratio, wherein the water production efficiency is the water production volume per unit power, and the energy consumption ratio is the water production volume per unit energy consumption;

[0037] A deep reinforcement learning model is constructed based on the state space, the action space and the reward function, wherein the deep reinforcement learning model includes an input layer, two hidden layers and an output layer; an experience replay buffer is set to store state-action transfer samples, and a greedy strategy is constructed for action exploration.

[0038] In an optional embodiment,

[0039] Dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework and construct training samples corresponding to extreme meteorological conditions through an adversarial sample generator, optimize the deep reinforcement learning model based on the training samples to obtain a candidate water production parameter control strategy set, execute the water production process based on the candidate water production parameter control strategy set and set the optimization target, including:

[0040] Constructing a meta-learning adaptive framework, calculating a network structure adjustment gradient based on the adaptive framework, and dynamically adjusting the depth, number of neurons, and connection mode of the deep reinforcement learning model according to the network structure adjustment gradient to obtain an improved deep reinforcement learning model;

[0041] Construct an adversarial sample generator, input the sample data under standard meteorological conditions into the adversarial sample generator, and generate training samples corresponding to extreme meteorological conditions of high temperature, low temperature, high humidity and low humidity through adversarial training;

[0042] The improved deep reinforcement learning model is optimized and trained based on the training samples using an online learning method, and the model parameters are updated by minimizing the expected value of the reward function to obtain a set of candidate water production parameter control strategies under extreme meteorological conditions.

[0043] In an optional embodiment,

[0044] The multi-objective Bayesian optimizer is used for joint optimization to generate a Pareto optimal solution set and select the optimal control parameters through a dynamic decision-making algorithm based on a confidence boundary. The update step size of the deep reinforcement learning model is dynamically adjusted based on the optimal control parameters to obtain the optimal water production strategy, including:

[0045] Construct a multi-objective Bayesian optimizer, take the water production efficiency objective function and the energy consumption ratio objective function as the optimization objectives of the multi-objective Bayesian optimizer, establish proxy models of the water production efficiency objective function and the energy consumption ratio objective function based on a Gaussian process regression model, use a radial basis function kernel to construct a kernel function of the proxy model, and perform sampling optimization on the proxy model based on an expected improvement criterion;

[0046] A Pareto front is constructed based on a set of sampling points obtained by the multi-objective Bayesian optimizer, a non-dominated solution is used as a Pareto optimal solution set corresponding to the Pareto front, and a confidence boundary dynamic decision algorithm is constructed. The confidence boundary dynamic decision algorithm calculates a confidence upper bound and a confidence lower bound of each solution in the Pareto optimal solution set based on a confidence parameter, the confidence parameter is dynamically adjusted with the number of iterations, and a solution with the smallest uncertainty in the Pareto optimal solution set is selected as an optimal control parameter based on the difference between the confidence upper bound and the confidence lower bound;

[0047] An adaptive step size adjustment function is designed based on the optimal control parameters, and the adaptive step size adjustment function is inversely proportional to the reward increment. The step size value and policy gradient obtained by the adaptive step size adjustment function are used to update the parameters of the deep reinforcement learning model. A momentum term is introduced to perform weighted averaging on the policy gradient. Based on the weighted policy gradient and step size, the parameters of the deep reinforcement learning model are updated in combination with the gradient correction enhancement mechanism and multi-scale parameter optimization to obtain the optimal water production strategy.

[0048] In an optional embodiment,

[0049] The step value and policy gradient obtained by the adaptive step adjustment function are used to update the parameters of the deep reinforcement learning model, a momentum term is introduced to perform weighted average on the policy gradient, and based on the weighted policy gradient and the step value, the parameters of the deep reinforcement learning model are updated in combination with the gradient correction enhancement mechanism and the multi-scale parameter optimization to obtain the optimal water production strategy, including:

[0050] A step size value is obtained based on an adaptive step size adjustment function, a periodic learning rate scheduling function is constructed, a periodic learning rate value is calculated based on a basic learning rate and a maximum learning rate, the periodic learning rate value is multiplied by the step size value and the inverse of a reward gain to obtain a combined learning rate, and in the initial stage of training, a ratio of the combined learning rate to a preset time step is used as a warm-up coefficient, and the combined learning rate is multiplied by the warm-up coefficient to obtain a warm-up learning rate;

[0051] Calculate the first-order moment estimate and the second-order moment estimate of the policy gradient, perform deviation correction on the first-order moment estimate and the second-order moment estimate based on a preset deviation correction coefficient to obtain a corrected first-order moment estimate and a corrected second-order moment estimate, multiply the warm-up learning rate by the ratio of the square roots of the corrected first-order moment estimate and the corrected second-order moment estimate, and add the ratio to the weight attenuation term to obtain a momentum term;

[0052] Dividing the parameters of the deep reinforcement learning model into fine-grained parameters, medium-grained parameters and coarse-grained parameters, calculating the update frequency of the parameters of different granularities based on the basic update frequency, calculating the policy gradients of the parameters of different granularities, and multiplying the policy gradients of the parameters of different granularities with corresponding weight coefficients and summing them to obtain a weighted policy gradient;

[0053] The weighted policy gradient is gradient clipped to obtain a clipped policy gradient, and the parameter update weight is calculated based on the learning rate weight, the weight of the momentum term and the decay weight. The product of the preheated learning rate and the clipped policy gradient, the momentum term and the original parameter are respectively multiplied by the corresponding parameter update weight and the sum is obtained to obtain the optimal water production strategy.

[0054] A second aspect of an embodiment of the present invention provides an intelligent prediction and regulation system for air-to-water production based on meteorological big data, comprising:

[0055] The first unit is used to collect meteorological monitoring data through meteorological sensors and perform standardization processing to generate a standardized meteorological feature matrix, extract periodic feature vectors and trend feature vectors through time series decomposition, and construct an enhanced feature space;

[0056] The second unit is used to build a transfer cross-domain learning network and add the pre-trained model knowledge in the fields of refrigeration and air conditioning and weather forecasting to the air-to-water prediction task through the knowledge transfer algorithm, and combine knowledge distillation to integrate expert knowledge from multiple source domains into the deep neural network;

[0057] The third unit is used to train a deep neural network based on the enhanced feature space and the back propagation algorithm and output the predicted water production volume hour by hour in the next day, use the predicted water production volume as the input of the state space and configure the action space, build a reward function according to the water production efficiency and energy consumption ratio, and build a deep reinforcement learning model based on the state space, action space and reward function;

[0058] The fourth unit is used to dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework and construct training samples corresponding to extreme meteorological conditions through an adversarial sample generator, optimize the deep reinforcement learning model based on the training samples to obtain a candidate water production parameter control strategy set, execute the water production process based on the candidate water production parameter control strategy set and set the optimization target;

[0059] The fifth unit is used to perform joint optimization through a multi-objective Bayesian optimizer, generate a Pareto optimal solution set, select optimal control parameters through a dynamic decision-making algorithm based on a confidence boundary, and dynamically adjust the update step size of the deep reinforcement learning model based on the optimal control parameters to obtain the optimal water production strategy.

[0060] In the present invention, by constructing a transfer cross-domain learning network, the pre-trained model knowledge in the field of refrigeration and air conditioning and the field of meteorological forecasting is migrated to the air-to-water prediction task, and the expert knowledge of multiple source domains is integrated with the knowledge distillation technology, which significantly improves the prediction accuracy and generalization ability of the model, and effectively solves the problems of data sparsity and difficulty in labeling in the air-to-water scenario. A meta-learning adaptive framework is used to dynamically adjust the network structure of the deep reinforcement learning model, and training samples corresponding to extreme meteorological conditions are constructed through an adversarial sample generator, which enhances the adaptability and robustness of the model in complex and changeable environments, improves the control accuracy and stability of the water production process, and performs joint optimization based on a multi-objective Bayesian optimizer and combines it with a dynamic decision-making algorithm based on a confidence boundary to achieve adaptive adjustment of water production parameters, which reduces energy consumption while ensuring water production efficiency, improves the overall operation efficiency of the system, and has good engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flow chart of an intelligent prediction and regulation method of air-to-water production based on meteorological big data according to an embodiment of the present invention;

[0062] Figure 2 This is a characteristic characterization quality assessment diagram of the air water production intelligent prediction and regulation method based on meteorological big data in an embodiment of the present invention;

[0063] Figure 3 This is a comparison chart of the adaptive learning rate adjustment effect of the air water production intelligent prediction and regulation method based on meteorological big data in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0066] Figure 1 FIG. 1 is a flow chart of an intelligent prediction and regulation method for air-to-water production based on meteorological big data according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0067] Meteorological monitoring data is collected through meteorological sensors and standardized to generate a standardized meteorological feature matrix. Periodic feature vectors and trend feature vectors are extracted through time series decomposition and an enhanced feature space is constructed.

[0068] Construct a transfer cross-domain learning network and add the pre-trained model knowledge in the fields of refrigeration and air conditioning and weather forecasting to the air-to-water prediction task through the knowledge transfer algorithm. Combine knowledge distillation to integrate expert knowledge from multiple source domains into the deep neural network.

[0069] Based on the enhanced feature space, a deep neural network is trained in combination with a back propagation algorithm to output the predicted water production volume hour by hour in the next day, the predicted water production volume is used as the input of the state space and the action space is configured, a reward function is constructed according to the water production efficiency and energy consumption ratio, and a deep reinforcement learning model is constructed based on the state space, the action space and the reward function;

[0070] Dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework and construct training samples corresponding to extreme meteorological conditions through an adversarial sample generator, optimize the deep reinforcement learning model based on the training samples to obtain a set of candidate water production parameter control strategies, execute the water production process based on the candidate water production parameter control strategy set and set the optimization target;

[0071] A multi-objective Bayesian optimizer is used for joint optimization to generate a Pareto optimal solution set, and the optimal control parameters are selected through a dynamic decision-making algorithm based on a confidence boundary. The update step size of the deep reinforcement learning model is dynamically adjusted based on the optimal control parameters to obtain the optimal water production strategy.

[0072] In an optional embodiment,

[0073] Meteorological monitoring data is collected through meteorological sensors and standardized to generate a standardized meteorological feature matrix. Periodic feature vectors and trend feature vectors are extracted through time series decomposition and an enhanced feature space is constructed, including:

[0074] Collect meteorological monitoring data through meteorological sensors, perform standardization on the meteorological monitoring data, and arrange the standardized meteorological monitoring data in time series to generate a standardized meteorological feature matrix;

[0075] Performing time series decomposition on the standardized meteorological feature matrix, constructing a trajectory matrix and performing singular value decomposition, and decomposing the standardized meteorological feature matrix into a periodic component and a trend component, extracting a periodic feature vector and a trend feature vector of the standardized meteorological feature matrix;

[0076] An enhanced feature space is constructed based on the periodic feature vector and the trend feature vector, wherein the enhanced feature space is used to characterize the periodic variation law and long-term variation trend of the meteorological monitoring data.

[0077] Meteorological monitoring data is collected through meteorological sensors. Meteorological sensors include temperature sensors, humidity sensors, air pressure sensors, wind speed sensors, etc. The collected meteorological monitoring data include meteorological elements such as temperature, humidity, air pressure, wind speed, etc. Taking temperature data as an example, temperature data for one year is collected with a sampling frequency of once per hour, and 8760 temperature values ​​are obtained.

[0078] Standardize the collected meteorological monitoring data. Standardization includes: normalizing the original data and mapping the data to the 0-1 interval; removing abnormal values ​​and correcting the data that exceeds the normal range; filling in missing values ​​and using interpolation methods to fill in missing data. Taking temperature data as an example, the temperature values ​​from -20℃ to 40℃ are mapped to the 0-1 interval, and the obviously abnormal temperature values ​​are removed. The missing temperature values ​​are supplemented by linear interpolation methods.

[0079] The standardized meteorological monitoring data are arranged in time series to generate a standardized meteorological feature matrix. Taking temperature data as an example, the standardized 8760 temperature values ​​are arranged in time sequence into an 8760×1 matrix.

[0080] The standardized meteorological feature matrix is ​​decomposed into time series. First, the trajectory matrix is ​​constructed, and the sliding window length is selected as 24 (corresponding to one day), and the original time series is reconstructed into a 365×24 trajectory matrix. The trajectory matrix is ​​subjected to singular value decomposition to obtain eigenvalues ​​and eigenvectors. According to the size of the eigenvalue, the main components are selected to reconstruct the time series to obtain the periodic component and the trend component.

[0081] Extract periodic eigenvectors and trend eigenvectors. Periodic eigenvectors reflect the periodic variation of data, such as daily and seasonal variations in temperature; trend eigenvectors reflect the long-term variation trend of data, such as the interannual variation trend of temperature.

[0082] An enhanced feature space is constructed based on periodic feature vectors and trend feature vectors. The periodic feature vectors and trend feature vectors are combined to construct a multidimensional feature space to characterize the periodic change law and long-term change trend of meteorological data.

[0083] In this embodiment, by standardizing and time-series decomposing the meteorological monitoring data, the periodic change patterns and long-term change trends in the data can be effectively extracted, thereby improving the accuracy and reliability of meteorological data analysis. The trajectory matrix reconstruction and singular value decomposition methods are used for time-series decomposition, which can effectively separate the periodic component and trend component in the meteorological data, providing a reliable data foundation for subsequent meteorological forecasting and analysis. The constructed enhanced feature space comprehensively considers the periodic changes and trend changes of meteorological data, and can more comprehensively characterize the changing characteristics of meteorological data, providing richer feature information for meteorological forecasting and analysis.

[0084] In an optional embodiment,

[0085] Construct a migration cross-domain learning network and add the pre-trained model knowledge in the refrigeration and air conditioning field and the weather forecast field to the air-to-water prediction task through the knowledge transfer algorithm. Combine knowledge distillation to integrate expert knowledge from multiple source domains into the deep neural network, including:

[0086] Constructing a transfer cross-domain learning network, the transfer cross-domain learning network includes a feature extractor, a task-specific classifier and a domain discriminator, the feature extractor adopts a convolutional neural network structure with residual connection, the domain discriminator adopts a three-layer fully connected network structure, and the feature extractor is used to extract feature vectors from the refrigeration and air-conditioning field and the weather forecast field;

[0087] Acquire feature statistical information based on the feature vector and input it into a kernel function generation network to obtain a kernel bandwidth parameter and a kernel function combination weight, identify the source domain of the feature vector based on a domain adversarial loss function to obtain a domain adversarial loss value, construct a maximum average difference metric based on the kernel bandwidth parameter and the kernel function combination weight, and calculate the difference between the feature vectors in the refrigeration and air conditioning field and the weather forecast field to obtain a maximum average difference value;

[0088] For each feature vector, an attention mechanism is used to calculate the importance of the pre-trained model knowledge in the refrigeration and air-conditioning field and the weather forecast field to obtain a corresponding attention weight, and the attention weight is multiplied by the feature vector and superimposed to obtain a fusion feature;

[0089] Based on the knowledge distillation loss function, the expert knowledge in the fusion feature is transferred to the deep neural network to obtain a knowledge distillation loss value, and the knowledge distillation loss value, the domain adversarial loss value and the maximum average difference value are weightedly combined to obtain an overall loss function;

[0090] An alternating optimization strategy is adopted to train the transfer cross-domain learning network, the domain discriminator parameters corresponding to the domain discriminator are fixed and the feature extractor parameters corresponding to the feature extractor are optimized based on the overall loss function, the feature extractor parameters are fixed and the domain discriminator parameters are optimized based on the output of the domain discriminator, and the parameter optimization process is iteratively performed through the back propagation algorithm until the overall loss function converges to obtain a trained transfer cross-domain learning network.

[0091] The structure of the feature extractor is designed in detail. The design of the convolution layer adopts a multi-branch structure. The main branch contains regular convolution operations, and the bypass branch is used to implement residual connections. The convolution operation in the main branch uses padding to keep the size of the feature map unchanged, and the convolution kernel parameters are initialized using a standard normal distribution. The residual connection is implemented through direct mapping or projection mapping to ensure that the output of the main branch matches the size of the bypass branch before adding them. After the convolution operation is completed, batch normalization is performed immediately to normalize the feature map in the channel dimension and introduce learnable scaling and translation parameters.

[0092] For the fully connected layer design of the domain discriminator, the number of neurons in the input layer corresponds to the output dimension of the feature extractor, the intermediate layer achieves feature compression by gradually reducing the dimension, and the output layer produces a two-dimensional discrimination result. The weight matrix of the fully connected layer is initialized with a uniform distribution, and the bias term is initialized to zero. A nonlinear activation function is introduced after each layer of linear transformation, and the rectified linear unit function is used to introduce nonlinear features.

[0093] After obtaining the feature vector, the design of the kernel function generation network adopts an adaptive kernel function structure. First, the multi-order statistical moment information of the feature vector is calculated, including the first-order moment (mean), second-order moment (variance), and higher-order moment (skewness, kurtosis). Based on these statistical information, a kernel function parameter generator is constructed, and the generator outputs the kernel bandwidth parameter and combination weight through a feedforward network structure. The kernel bandwidth parameter determines the shape of the Gaussian kernel function, and the combination weight is used for the linear combination of multiple basic kernel functions.

[0094] The calculation process of domain adversarial loss is divided into two steps: input the feature vector into the domain discriminator to obtain a two-dimensional output, normalize it to obtain the domain discrimination probability, and calculate the cross entropy between the probability value and the true domain label. The maximum average difference is calculated based on the generated kernel function parameters by calculating the distance between the two domains in the reproducing kernel Hilbert space.

[0095] The implementation of the attention mechanism uses a scaled dot product attention structure. The feature vector is converted into a query vector, a key vector, and a value vector respectively. The similarity score is calculated by the dot product of the query vector and the key vector. The attention weight is obtained after scaling and normalization. The attention weight is multiplied by the value vector to obtain the weighted feature. Multi-head attention is achieved by parallel calculation of multiple groups of attention results and splicing them. Finally, the weighted features are summed to obtain the fused feature vector.

[0096] During the knowledge distillation process, the fused features of the pre-trained model are used as soft labels. The smoothness of the soft labels is adjusted by the temperature parameter. The larger the temperature value, the smoother the distribution. The distillation loss is calculated by calculating the relative entropy of the soft labels and the output of the student network. The overall loss function combines the three loss terms through adjustable weight coefficients, which can be adjusted according to the actual task requirements.

[0097] In terms of optimization strategy, the gradient descent method is used to update the network parameters. For the optimization of the feature extractor, the gradient of the overall loss function with respect to the convolution kernel and batch normalization parameters is calculated, and the parameters are updated in the opposite direction of the gradient. For the optimization of the domain discriminator, only the domain discrimination loss is considered, and the gradient of the fully connected layer parameters is calculated for update. The loss function value is evaluated after each update, and the historical optimal model parameters are recorded. When the loss change of multiple consecutive iterations is less than the threshold, the model is considered to have converged.

[0098] To ensure training stability, an early stopping strategy and learning rate scheduling mechanism are introduced. The early stopping strategy determines whether to terminate training early based on the performance of the validation set to avoid overfitting. The learning rate scheduling dynamically adjusts the learning rate size according to the training rounds. A larger learning rate is used for rapid convergence in the early stage, and a smaller learning rate is used for fine adjustment in the later stage. At the same time, the optimal model parameters during the training process are saved as the final training results.

[0099] For example, in the field of refrigeration and air conditioning of air-to-water systems, the original data includes refrigeration system operation data, including: compressor inlet and outlet temperature, pressure, flow rate, condenser inlet and outlet temperature, pressure, flow rate, evaporator inlet and outlet temperature, pressure, flow rate, expansion valve opening and other operating parameters. In the field of weather forecasting, the original data includes ambient temperature, relative humidity, atmospheric pressure, wind speed, dew point temperature and other meteorological element data.

[0100] First, the data in the field of refrigeration and air conditioning are input into the feature extractor, and the operating characteristics of the refrigeration system are extracted through multi-layer convolution processing. These characteristics reflect the working relationship and energy conversion law between the various components of the refrigeration system. Similarly, the data in the field of meteorological forecast are input into the same feature extractor to extract the change characteristics between meteorological elements. These characteristics reflect the dynamic change law of the atmospheric environment.

[0101] The extracted feature vector is input into the kernel function generation network, and the network generates kernel bandwidth parameters and combination weights based on the statistical information of the features. For example, when the compressor data of the refrigeration system is input, the network will pay more attention to the correlation characteristics between the compressor and other components; when the ambient temperature data is input, the network will pay more attention to the coupling characteristics of temperature and humidity.

[0102] The domain discriminator identifies the source of the feature vector and determines whether the feature comes from the refrigeration and air conditioning field or the weather forecast field. The maximum average difference is used to calculate the difference between the features of the two fields and to evaluate the domain adaptability of the feature extractor.

[0103] The attention mechanism evaluates the importance of knowledge in different fields. For example, when predicting water production, the evaporation temperature of the refrigeration system and the ambient dew point temperature are both important parameters. The attention mechanism will dynamically assign higher weights to these key features based on the current working conditions. When the ambient humidity is low, the attention will be paid more attention to the working conditions of the refrigeration system compressor; when the ambient humidity is high, more attention will be paid to changes in ambient temperature.

[0104] The knowledge distillation process transfers the fused expert knowledge to the target network. For example, the optimal control strategy of the refrigeration system under different environmental conditions and the influence of meteorological factors on water production efficiency will be transferred to the target network through knowledge distillation.

[0105] In the alternating optimization process, the feature extractor is first optimized to extract the common features of the refrigeration system and the meteorological environment, and then the domain discriminator is optimized to enhance its ability to distinguish features from different fields. For example, the feature extractor gradually learns the correlation between compressor power and ambient temperature, and the domain discriminator learns to determine which field this correlation belongs to.

[0106] The trained network can make comprehensive use of knowledge in the fields of refrigeration and air conditioning and weather forecasting to achieve accurate water production forecasting. For example, when new environmental conditions and refrigeration system parameters are input, the network can predict the optimal water production conditions and the corresponding water production output. When environmental conditions change, the network can adaptively adjust the control strategy to maintain a high water production efficiency.

[0107] In this embodiment, the pre-trained model knowledge in the fields of refrigeration and air conditioning and weather forecasting is effectively utilized through transfer learning and knowledge distillation technology, the performance of the air-to-water prediction task is improved, the attention mechanism and kernel function adaptive selection strategy are adopted to realize the dynamic fusion of knowledge from different source domains, and enhance the adaptability of the model to different working conditions. Based on domain adversarial training and maximum average difference metric, the feature distribution difference between the source domain and the target domain is reduced, and the generalization error of the model in the target domain is reduced;

[0108] In the existing air-to-water system, the prediction of water production usually adopts a single-domain knowledge modeling method, which only builds a prediction model based on refrigeration and air-conditioning knowledge, takes the operating parameters of various components of the refrigeration system as input for modeling, and ignores the impact of meteorological environmental changes on water production efficiency, or only relies on meteorological forecasting knowledge to predict theoretical water production based on environmental parameters, without fully considering the actual operating characteristics of the refrigeration system. The single-domain modeling method is difficult to accurately describe the complex coupling relationship between the refrigeration system and the environment during the air-to-water process, resulting in insufficient prediction accuracy, especially in extreme weather or non-rated equipment conditions, the prediction error is large;

[0109] This embodiment extracts deep features in the fields of refrigeration and air conditioning and weather forecasting through a convolutional neural network with residual connections. The design of residual connections effectively alleviates the gradient vanishing problem of deep networks and improves the accuracy of feature extraction. The introduction of domain discriminators and maximum mean difference metrics realizes the alignment of feature distributions in different fields, making the extracted features have better domain adaptability. Knowledge distillation technology efficiently transfers expert knowledge in the two fields to the target prediction model, making full use of the prior knowledge contained in the pre-trained model. In summary, this embodiment significantly improves the accuracy, generalization ability and environmental adaptability of the prediction model, and provides strong support for the intelligent control of the air-to-water system.

[0110] In an optional embodiment,

[0111] Acquiring feature statistical information based on the feature vector and inputting it into a kernel function generation network to obtain a kernel bandwidth parameter and a kernel function combination weight, performing source domain identification on the feature vector based on a domain adversarial loss function to obtain a domain adversarial loss value, constructing a maximum average difference metric based on the kernel bandwidth parameter and the kernel function combination weight, and calculating the feature vector difference between the refrigeration and air conditioning field and the weather forecast field to obtain the maximum average difference value includes:

[0112] Based on the feature vector, a feature pyramid structure is constructed in multiple network layers of a feature extractor, wherein the feature pyramid structure includes a shallow feature vector, a middle feature vector and a deep feature vector, wherein the shallow feature vector includes local texture information, the middle feature vector includes local semantic information, and the deep feature vector includes global semantic information;

[0113] Calculating the mutual information between the shallow feature vector, the middle feature vector and the deep feature vector to obtain an inter-layer correlation score, and selecting the feature level combination with the strongest discriminability based on the inter-layer correlation score to obtain a combined feature vector;

[0114] Acquire the mean and variance of the combined feature vector to construct feature statistical information, input the feature statistical information into a kernel function generation network, and the kernel function generation network outputs a kernel bandwidth parameter and a kernel function combination weight;

[0115] Calculating the inter-class distance and intra-class distance of the combined feature vector in the refrigeration and air-conditioning field to obtain the feature distance in the refrigeration and air-conditioning field, and calculating the feature variance of the combined feature vector in the weather forecast field to obtain the feature variance in the weather forecast field;

[0116] Calculating the feature dimension importance based on the feature distance in the refrigeration and air conditioning field and the feature variance in the weather forecast field, and weighting the feature dimension importance with the combined feature vector to obtain a weighted feature vector;

[0117] The weighted feature vector is input into a domain discriminator, and the source domain of the weighted feature vector is identified based on a domain adversarial loss function to obtain a domain adversarial loss value. A maximum average difference metric is constructed based on the kernel bandwidth parameter and the kernel function combination weight. The maximum average difference metric is used to calculate the difference between the weighted feature vectors in the refrigeration and air-conditioning field and the meteorological forecast field to obtain a maximum average difference value.

[0118] The feature pyramid structure is constructed at different network layers of the feature extractor. In the shallow network, local texture information is extracted through convolution operations with a smaller receptive field, and the underlying features such as edges and textures in the data are captured to form shallow feature vectors. In the middle layer of the network, as the receptive field expands, the semantic information of the local area is extracted to obtain a more representative middle-layer feature vector. In the deep network, the receptive field range is expanded through multiple convolution and downsampling operations to extract deep feature vectors with global semantic information.

[0119] For the extracted multi-level feature vectors, it is necessary to evaluate the correlation between features at different levels. By calculating the mutual information between shallow and middle layers, middle and deep layers, and shallow and deep feature vectors, the information redundancy and complementarity between feature levels are quantified. Based on the calculated inter-layer correlation scores, the feature level combination with the strongest discriminative ability is selected, and the feature vectors of the selected levels are fused to obtain the combined feature vector.

[0120] For the combined feature vector, its distribution characteristics in the feature space are calculated. First, the mean and variance of the feature vector in each dimension are obtained to construct feature statistics. This statistical information is input into a pre-designed kernel function generation network, which adaptively generates kernel bandwidth parameters and kernel function combination weights through multi-layer transformations.

[0121] In the field of refrigeration and air conditioning, the inter-class distance and intra-class distance of the combined feature vector are calculated. The inter-class distance reflects the discrimination of sample features of different categories, and the intra-class distance reflects the compactness of sample features of the same category. The two together constitute the characteristic distance in the field of refrigeration and air conditioning. For the field of weather forecasting, the variance distribution of the combined feature vector in each dimension is calculated to obtain the characteristic variance in the field of weather forecasting.

[0122] Based on the calculated feature distance in the refrigeration and air conditioning field and feature variance in the weather forecast field, the importance of each dimension in the feature space is evaluated. Feature dimensions with strong discriminability and large variance are given higher importance scores. The obtained feature dimension importance is weighted with the combined feature vector to obtain a weighted feature vector.

[0123] The weighted feature vector is input into the domain discriminator for source domain identification. The domain discrimination loss value of the feature vector is calculated through the domain adversarial loss function, and the maximum average difference metric is constructed based on the parameters of the kernel function generation network output. This metric is used to calculate the distribution difference between the weighted feature vectors in the refrigeration and air conditioning field and the meteorological forecast field, and the maximum average difference value is obtained.

[0124] For example, in the air-to-water system, the input data in the field of refrigeration and air conditioning includes the operating parameters of components such as compressors, condensers, and evaporators, and the input data in the field of weather forecasting includes environmental parameters such as temperature, humidity, and air pressure. These data are first processed by the feature extractor in multiple layers:

[0125] The shallow network extracts the basic change trends of equipment operating parameters and the fluctuation characteristics of environmental parameters; the middle network captures the correlation between the components of the refrigeration system and the coupling characteristics between meteorological elements; the deep network extracts the overall operation laws of the refrigeration cycle and the macroscopic change characteristics of the atmospheric environment.

[0126] By calculating the mutual information between feature levels, it is found that shallow features can better characterize the instantaneous changes of parameters, middle-level features reflect the mutual influence between parameters, and deep features reflect the overall operating status of the system. Based on the complementarity of these features, the most discriminative layer combination is selected to construct a combined feature vector.

[0127] In the field of refrigeration and air conditioning, the characteristic distances under different operating conditions are calculated, and the characteristics of high-efficiency conditions and low-efficiency conditions show large inter-class differences. In the field of meteorological forecasting, the characteristic variance distribution under different weather conditions reflects the changing laws of environmental parameters.

[0128] Based on the characteristic distance and variance analysis, the key operating parameters of the refrigeration system and the dominant meteorological factors are given higher weights. For example, in the field of refrigeration and air conditioning, parameters such as compressor power and evaporation temperature are given higher weights, such as 0.3 and 0.35; in the field of meteorological forecasting, parameters such as relative humidity and dew point temperature are given higher weights, such as 0.35 and 0.42.

[0129] The weighted feature vector is input into the domain discriminator for recognition. At the same time, the maximum average difference metric is used to evaluate the difference between the features of the two domains, providing a basis for subsequent knowledge transfer and fusion.

[0130] Figure 2 : is a characteristic characterization quality assessment diagram of the intelligent prediction and regulation method of air water production based on meteorological big data in an embodiment of the present invention, such as Figure 2 As shown in the figure, the feature representation quality evaluation diagram shows the comparison between the present technical solution and PCA and AutoEncoder in the feature dimension importance evaluation. The horizontal axis in the figure is the feature dimension index (1-50), and the vertical axis is the dimension importance score (0-1). In the key feature interval of dimension index 15-25, the importance score of the present technical solution is stably maintained between 0.82-0.88, which is significantly higher than PCA (0.62-0.68) and AutoEncoder (0.55-0.61), especially in the dimension 20. At , this technical solution achieved the highest score of 0.87, while PCA and AutoEncoder were 0.67 and 0.59 respectively. In the secondary feature interval of dimension 30-40, this technical solution showed a more delicate discrimination ability, and the importance score showed a reasonable fluctuation range (0.45-0.65), while the scores of other methods in this interval were relatively flat (PCA: 0.52-0.58, AutoEncoder: 0.42-0.48). Therefore, this technical solution has significant advantages over traditional technical means;

[0131] In this embodiment, through the feature pyramid structure and inter-layer correlation analysis, multi-scale feature information can be effectively extracted, the feature expression ability can be enhanced, and the performance of transfer learning can be improved. The kernel function is used to generate the network adaptive learning kernel function parameters, which avoids the limitations of manually setting parameters and makes the measurement method more flexible and accurate. The combination of domain adversarial learning and maximum average difference measurement not only ensures the discriminability of features but also reduces the difference between domains, achieving better cross-domain knowledge transfer effect;

[0132] The existing feature extraction methods of air-to-water systems usually adopt a single-level feature extraction strategy, which cannot simultaneously capture the multi-scale feature information of the data, resulting in limited feature expression capabilities. In the feature selection process, a fixed weight method is often used, which makes it difficult to dynamically adjust the importance of features according to actual working conditions. At the same time, there is a lack of effective domain adaptation mechanism when processing data in different fields, resulting in a large domain offset of the extracted features, which affects the generalization ability of the model. This embodiment constructs a feature pyramid structure in different network layers of the feature extractor to achieve multi-level feature extraction from local texture to global semantics, evaluates the complementarity of features at different levels by calculating inter-layer mutual information, selects the most discriminative feature combination, and effectively improves the feature expression capability. By calculating the inter-class distance and intra-class distance in the field of refrigeration and air conditioning, and the feature variance in the field of meteorological forecasting, the adaptive calculation of feature importance is achieved. In summary, this embodiment achieves efficient feature extraction and domain adaptation of data in the fields of refrigeration and air conditioning and meteorological forecasting, significantly improves the feature expression capability, discrimination capability and generalization capability, and provides more reliable feature support for performance prediction and optimization control of air-to-water systems, which is of great significance to improving the overall performance of the system.

[0133] In an optional embodiment,

[0134] Based on the enhanced feature space, the deep neural network is trained in combination with the back propagation algorithm and the predicted water production volume is output hour by hour in the next day. The predicted water production volume is used as the input of the state space and the action space is configured. The reward function is constructed according to the water production efficiency and energy consumption ratio. The deep reinforcement learning model is constructed based on the state space, action space and reward function, including:

[0135] A deep neural network is constructed based on the enhanced feature space, wherein the deep neural network includes an input layer, three hidden layers and an output layer, wherein the hidden layer uses a rectified linear unit activation function; a back propagation algorithm is used to train the deep neural network to obtain hourly predicted water production in the next day;

[0136] The predicted water production volume is constructed as the input of the state space, and an action space including compressor frequency, fan speed and system start / stop state is configured; a reward function is constructed based on water production efficiency and energy consumption ratio, wherein the water production efficiency is the water production volume per unit power, and the energy consumption ratio is the water production volume per unit energy consumption;

[0137] A deep reinforcement learning model is constructed based on the state space, the action space and the reward function, wherein the deep reinforcement learning model includes an input layer, two hidden layers and an output layer; an experience replay buffer is set to store state-action transfer samples, and a greedy strategy is constructed for action exploration.

[0138] A deep neural network is constructed to predict water production. The network consists of an input layer, three hidden layers, and an output layer. The input layer receives feature data in the enhanced feature space, including historical water production, ambient temperature, humidity, and other data. The number of neurons in the three hidden layers is set to 128, 64, and 32, respectively, and rectified linear units are used as activation functions. The number of neurons in the output layer is 24, corresponding to the predicted water production in the next 24 hours. The network is trained using the back propagation algorithm, the learning rate is set to 0.001, the number of training rounds is 1000, and 80% of the samples are randomly selected for training in each round.

[0139] The predicted 24-hour water production is used as the state space input, and the action space is configured to include the compressor frequency (range 20-60Hz, step size 5Hz), fan speed (range 500-1500rpm, step size 100rpm) and system start and stop status (0 for shutdown, 1 for operation). The water production efficiency is defined based on the water production under unit power, and the unit is kg / kW; the energy consumption ratio is defined based on the water production under unit energy consumption, and the unit is kg / kWh. The reward function is constructed by weighted summing the water production efficiency and energy consumption ratio, with weights of 0.6 and 0.4 respectively.

[0140] A deep reinforcement learning model was constructed, which included an input layer, two hidden layers (with 64 and 32 neurons respectively) and an output layer. The experience replay buffer size was set to 10,000 to store state, action, reward and next state transition samples. The ε-greedy strategy was used for action exploration, with an initial ε value of 0.9, which decreased to 0.01 at a decay rate of 0.995. 256 samples were randomly sampled for update in each round of training, and the number of training rounds was 5,000.

[0141] In actual applications, the characteristic data at the current moment is input, and the predicted water production value for the next 24 hours is obtained through the prediction network. The predicted water production is input into the reinforcement learning model, and the model outputs the optimal compressor frequency, fan speed, and system start and stop status. The system performs corresponding actions and records the actual operation data for subsequent online update and optimization of the model.

[0142] In this embodiment, a deep neural network is used to accurately predict the future water production, providing a reliable decision-making basis for system control. Deep reinforcement learning is combined to achieve intelligent optimization of system control parameters. Experience replay and greedy strategies are used to ensure the stability and exploration efficiency of model training, so that the model has good generalization and adaptability, and can sustainably optimize the system operation effect.

[0143] In an optional embodiment,

[0144] Dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework and construct training samples corresponding to extreme meteorological conditions through an adversarial sample generator, optimize the deep reinforcement learning model based on the training samples to obtain a candidate water production parameter control strategy set, execute the water production process based on the candidate water production parameter control strategy set and set the optimization target, including:

[0145] Constructing a meta-learning adaptive framework, calculating a network structure adjustment gradient based on the adaptive framework, and dynamically adjusting the depth, number of neurons, and connection mode of the deep reinforcement learning model according to the network structure adjustment gradient to obtain an improved deep reinforcement learning model;

[0146] Construct an adversarial sample generator, input the sample data under standard meteorological conditions into the adversarial sample generator, and generate training samples corresponding to extreme meteorological conditions of high temperature, low temperature, high humidity and low humidity through adversarial training;

[0147] The improved deep reinforcement learning model is optimized and trained based on the training samples using an online learning method, and the model parameters are updated by minimizing the expected value of the reward function to obtain a set of candidate water production parameter control strategies under extreme meteorological conditions.

[0148] Construct a meta-learning adaptive framework, which includes a network structure search module and a parameter optimization module. The network structure search module uses a gradient-based method to guide the structural adjustment by calculating the gradient of the loss on the validation set to the network structure parameters. In the specific implementation, the depth range is set to 3-10 layers, the number of neurons in each layer ranges from 32-256, and the connection methods include residual connection and dense connection. According to the gradient update direction, the network depth, number of neurons and connection method are gradually adjusted, and the structural optimization is completed when the performance of the validation set no longer improves.

[0149] Construct an adversarial sample generator, which includes two networks: the generator and the discriminator. The generator adopts an encoder-decoder structure, and the input is data such as temperature and humidity under standard meteorological conditions. It generates extreme meteorological condition samples through multi-layer convolution and deconvolution networks. The discriminator is used to judge the authenticity of the generated samples. During training, meteorological data for 7 consecutive days under standard conditions are collected as a training set, including 5,000 data samples with a temperature of 20-25°C and a relative humidity of 50-70%. Through adversarial training, 1,000 training samples of extreme conditions such as high temperature (35-40°C), low temperature (0-5°C), high humidity (85-95%) and low humidity (20-30%) are generated.

[0150] The deep reinforcement learning model is optimized by online learning. The state space is set to include 12 parameters such as influent water quality and meteorological conditions, and the action space includes 4 control parameters such as membrane flux and operating pressure. The reward function comprehensively considers three aspects: water production, energy consumption and membrane life. Every time a new set of extreme working condition training samples is obtained, the model parameters are updated immediately. After 1,000 rounds of iterative training, the average reward value of the model under various extreme working conditions has increased by more than 30%, and the corresponding optimal control strategy set has been obtained.

[0151] In this embodiment, the network structure is dynamically adjusted through a meta-learning adaptive framework, which improves the generalization ability and robustness of the model under different working conditions. The extreme weather conditions training samples constructed based on the adversarial sample generator effectively solve the problem of scarcity of extreme conditions data in actual operation, enabling the model to learn control strategies for extreme conditions in advance, avoiding the limitations of traditional methods that rely on manual experience, and continuously optimizing the model through online learning, realizing real-time updating and dynamic adjustment of control strategies, reducing energy consumption while ensuring water production, and significantly improving the overall operating efficiency and economy of the system.

[0152] In an optional embodiment,

[0153] The multi-objective Bayesian optimizer is used for joint optimization to generate a Pareto optimal solution set and select the optimal control parameters through a dynamic decision-making algorithm based on a confidence boundary. The update step size of the deep reinforcement learning model is dynamically adjusted based on the optimal control parameters to obtain the optimal water production strategy, including:

[0154] Construct a multi-objective Bayesian optimizer, take the water production efficiency objective function and the energy consumption ratio objective function as the optimization objectives of the multi-objective Bayesian optimizer, establish proxy models of the water production efficiency objective function and the energy consumption ratio objective function based on a Gaussian process regression model, use a radial basis function kernel to construct a kernel function of the proxy model, and perform sampling optimization on the proxy model based on an expected improvement criterion;

[0155] A Pareto front is constructed based on a set of sampling points obtained by the multi-objective Bayesian optimizer, a non-dominated solution is used as a Pareto optimal solution set corresponding to the Pareto front, and a confidence boundary dynamic decision algorithm is constructed. The confidence boundary dynamic decision algorithm calculates a confidence upper bound and a confidence lower bound of each solution in the Pareto optimal solution set based on a confidence parameter, the confidence parameter is dynamically adjusted with the number of iterations, and a solution with the smallest uncertainty in the Pareto optimal solution set is selected as an optimal control parameter based on the difference between the confidence upper bound and the confidence lower bound;

[0156] An adaptive step size adjustment function is designed based on the optimal control parameters, and the adaptive step size adjustment function is inversely proportional to the reward increment. The step size value and policy gradient obtained by the adaptive step size adjustment function are used to update the parameters of the deep reinforcement learning model. A momentum term is introduced to perform weighted averaging on the policy gradient. Based on the weighted policy gradient and step size, the parameters of the deep reinforcement learning model are updated in combination with the gradient correction enhancement mechanism and multi-scale parameter optimization to obtain the optimal water production strategy.

[0157] A multi-objective Bayesian optimizer is constructed. Water production efficiency and energy consumption ratio are taken as two optimization objectives, and their corresponding proxy models are established respectively. The proxy model is implemented based on Gaussian process regression, and the correlation between sample points is described by the radial basis function kernel. For each proxy model, the next most valuable sampling point is determined based on the expected improvement criterion, which considers both the predicted mean and variance to achieve a balance between exploration and utilization.

[0158] After obtaining the set of sampling points, the Pareto front is constructed. All sampling points are sorted non-dominated, and the solutions that are not dominated by any other solution are taken as the Pareto optimal solution set. A confidence boundary dynamic decision algorithm is constructed, which introduces a confidence parameter to balance the stability and adaptability of the decision. For each solution in the Pareto optimal solution set, its confidence upper and lower bounds are calculated based on the confidence parameter. The confidence parameter will be dynamically adjusted with the iteration process. It is larger in the early stage of optimization to encourage exploration, and gradually decreases in the later stage to promote convergence. The uncertainty of each solution is evaluated by calculating the difference of the confidence boundary, and the solution with the smallest uncertainty is selected as the optimal control parameter.

[0159] Based on the optimal control parameters obtained, an adaptive step size adjustment function is designed. This function is inversely proportional to the reward increment obtained by the system. When the reward increment is large, a smaller step size is used to ensure stability, and when the reward increment is small, a larger step size is used to accelerate convergence. The calculated step size value is combined with the policy gradient to update the parameters of the deep reinforcement learning model. At the same time, the momentum term is introduced to perform weighted averaging of the historical policy gradients to reduce the volatility of parameter updates.

[0160] In the parameter update process, the gradient correction enhancement mechanism is combined to normalize the gradient and introduce an adaptive learning rate. Through the multi-scale parameter optimization strategy, the model parameters are optimized at different scales to improve the optimization efficiency. Finally, the optimal water production strategy that can balance water production efficiency and energy consumption is obtained.

[0161] For example, in the optimization control of the air-to-water system, the water production efficiency objective function considers the water production per unit time, and the energy consumption ratio objective function considers the water production per unit energy consumption. The multi-objective Bayesian optimizer first samples different combinations of control parameters such as compressor frequency and fan speed under the initial working conditions.

[0162] The surrogate model is established through Gaussian process regression to predict the water production efficiency and energy consumption ratio under different control parameter combinations. The radial basis function kernel is used to calculate the correlation between sample points, and new sampling points are selected based on the expected improvement criterion. For example, when the water production efficiency predicted by a certain operating point is high and the prediction uncertainty is large, this point will be selected as the next sampling point.

[0163] The Pareto frontier is constructed based on the sampling results. For example, some operating points cannot be dominated by other points in terms of water production efficiency and energy consumption ratio. These points constitute the Pareto optimal solution set. The uncertainty of these solutions is evaluated through the confidence boundary dynamic decision algorithm. For example, in the initial stage, the confidence parameter is large, which allows a large exploration space; as the optimization process progresses, the confidence parameter gradually decreases, and more attention is paid to known high-quality solutions.

[0164] For the selected optimal control parameters, an adaptive step size adjustment function is designed. When the system achieves significant performance improvement under a certain condition, a smaller step size is used for fine-tuning; when the performance improvement is not obvious, a larger step size is used to try a new control strategy. After the momentum term is introduced, the update of the control parameters will take into account the historical gradient information to avoid falling into the local optimum.

[0165] Through gradient correction and multi-scale optimization, the control strategy is optimized at different time scales. For example, the rapid response of the compressor frequency is optimized on a short time scale, and the switching strategy of the system operation mode is optimized on a long time scale. The final water production strategy can adaptively adjust the control parameters according to environmental conditions and system status to achieve a dynamic balance between water production efficiency and energy consumption.

[0166] In this embodiment, a method combining multi-objective Bayesian optimization and Gaussian process regression is adopted to efficiently construct a proxy model of the objective function, significantly reduce the number of sampling times, accelerate the optimization convergence speed, and achieve a good balance between exploration and utilization by adaptively adjusting the confidence parameters based on the dynamic decision algorithm of the confidence boundary, effectively avoiding falling into the local optimal solution. The introduction of adaptive step size adjustment and momentum term mechanism, combined with gradient correction and multi-scale optimization, significantly improves the training stability and convergence performance of the deep reinforcement learning model, ensuring the optimal water production strategy.

[0167] The existing control optimization method of the air-to-water system is difficult to take into account both the water production efficiency and the energy consumption ratio at the same time. It often leads to a significant increase in energy consumption while improving the water production efficiency. Conventional multi-objective optimization algorithms are prone to fall into local optimality when dealing with high-dimensional nonlinear optimization problems, and the sampling efficiency is low. The existing reinforcement learning method uses a fixed step size in the parameter update process, which is difficult to adapt to the dynamic characteristics of the system and lacks effective guarantee for the stability of the gradient update. This embodiment constructs a multi-objective Bayesian optimizer, takes the water production efficiency and the energy consumption ratio as optimization objectives, and uses the Gaussian process regression model to establish a proxy model of the objective function. The basis function kernel effectively captures the nonlinear relationship in the parameter space, and uses the expected improvement criterion to achieve efficient sampling optimization. By constructing the Pareto front, all non-dominated solutions are retained, avoiding the subjectivity of artificial weighting of multiple objectives in traditional methods. In summary, this embodiment achieves comprehensive optimization of the control strategy of the air-to-water system through an innovative combination of multi-objective Bayesian optimization, dynamic decision-making mechanism and adaptive deep reinforcement learning, and has achieved significant improvements in water production efficiency and energy consumption balance, decision reliability, training stability and optimization efficiency, providing a more efficient and reliable solution for the intelligent control of the air-to-water system.

[0168] In an optional embodiment,

[0169] The step value and policy gradient obtained by the adaptive step adjustment function are used to update the parameters of the deep reinforcement learning model, a momentum term is introduced to perform weighted average on the policy gradient, and based on the weighted policy gradient and the step value, the parameters of the deep reinforcement learning model are updated in combination with the gradient correction enhancement mechanism and the multi-scale parameter optimization to obtain the optimal water production strategy, including:

[0170] A step size value is obtained based on an adaptive step size adjustment function, a periodic learning rate scheduling function is constructed, a periodic learning rate value is calculated based on a basic learning rate and a maximum learning rate, the periodic learning rate value is multiplied by the step size value and the inverse of a reward gain to obtain a combined learning rate, and in the initial stage of training, a ratio of the combined learning rate to a preset time step is used as a warm-up coefficient, and the combined learning rate is multiplied by the warm-up coefficient to obtain a warm-up learning rate;

[0171] Calculate the first-order moment estimate and the second-order moment estimate of the policy gradient, perform deviation correction on the first-order moment estimate and the second-order moment estimate based on a preset deviation correction coefficient to obtain a corrected first-order moment estimate and a corrected second-order moment estimate, multiply the warm-up learning rate by the ratio of the square roots of the corrected first-order moment estimate and the corrected second-order moment estimate, and add the ratio to the weight attenuation term to obtain a momentum term;

[0172] Dividing the parameters of the deep reinforcement learning model into fine-grained parameters, medium-grained parameters and coarse-grained parameters, calculating the update frequency of the parameters of different granularities based on the basic update frequency, calculating the policy gradients of the parameters of different granularities, and multiplying the policy gradients of the parameters of different granularities with corresponding weight coefficients and summing them to obtain a weighted policy gradient;

[0173] The weighted policy gradient is gradient clipped to obtain a clipped policy gradient, and the parameter update weight is calculated based on the learning rate weight, the weight of the momentum term and the decay weight. The product of the preheated learning rate and the clipped policy gradient, the momentum term and the original parameter are respectively multiplied by the corresponding parameter update weight and the sum is obtained to obtain the optimal water production strategy.

[0174] The adaptive step size adjustment function calculates the step size value according to the system operation status and performance indicators. The periodic learning rate scheduling function calculates the periodic learning rate value according to the training cycle change law by setting the basic learning rate and maximum learning rate parameters. The calculated periodic learning rate value and step size value are multiplied by the inverse of the current reward gain of the system to obtain a combined learning rate that can be adaptively adjusted. In the initial stage of deep reinforcement learning model training, in order to ensure training stability, the combined learning rate is ratioed with the preset time step to obtain the preheating coefficient. The combined learning rate is multiplied by the preheating coefficient to obtain the preheating learning rate, so as to achieve a smooth adjustment of the learning rate.

[0175] The policy gradient of the deep reinforcement learning model is moment estimated, and the first-order moment estimate and the second-order moment estimate are calculated. In order to eliminate the accumulated bias in the estimation process, a preset bias correction coefficient is introduced, and the first-order moment estimate and the second-order moment estimate are bias corrected respectively to obtain the corrected first-order moment estimate and the corrected second-order moment estimate. The warm-up learning rate is multiplied by the corrected first-order moment estimate, divided by the square root of the corrected second-order moment estimate, and added to the weight decay term to obtain the momentum term for parameter update, which can smooth the parameter update process and improve training stability.

[0176] The parameters in the deep reinforcement learning model are divided into three levels according to the update requirements and sensitivity: fine-grained parameters are responsible for fast response and fine adjustment, medium-grained parameters handle changes in medium time scales, and coarse-grained parameters control long-term policy adjustments. Based on the set basic update frequency, the update frequency of the three granularity parameters is calculated respectively, and the update frequency decreases as the parameter granularity increases. The policy gradients are calculated for the parameters of the three granularities respectively, and the policy gradient reflects the degree of influence of the parameters on the system performance. The policy gradients of the parameters of different granularities are multiplied and summed with their corresponding weight coefficients to obtain the weighted policy gradient that comprehensively considers multiple time scales.

[0177] The weighted policy gradient is subjected to gradient clipping. By setting the gradient norm threshold, the excessive gradient value is limited to a reasonable range to obtain the clipped policy gradient, thus avoiding the gradient explosion problem during the parameter update process. Based on the system performance indicators and training progress, the learning rate weight, momentum term weight and decay weight are calculated. These weights are used to balance the contribution of different update terms to obtain the parameter update weight. The product term of the warm-up learning rate and the clipped policy gradient, the momentum term containing historical information, and the original parameter reflecting the parameter inertia are multiplied by their corresponding parameter update weights, and the results are summed to obtain the final optimal water production strategy.

[0178] Exemplarily, during the operation of the air-to-water system, the system initially sets the adaptive step value to 0.01. When it is detected that the system reward gain reaches 2.5, the adaptive step adjustment function automatically adjusts the step value to 0.004. The periodic learning rate scheduling function is based on the set basic learning rate of 0.001 and the maximum learning rate of 0.01. It is adjusted within a training cycle of 1000 steps. When the number of training steps reaches 250 steps, the periodic learning rate value is 0.0055. The combined learning rate calculated at this time is 0.004×0.0055×0.4=0.0000088. Since it is in the early stage of training, the preset time step of 100 is used for preheating. The preheating coefficient is 0.0000088÷100=8.8e-8, and the preheating learning rate is 0.0000088×8.8e-8=7.744e-13.

[0179] The first-order moment estimate of the policy gradient is calculated to be 0.15, and the second-order moment estimate is 0.023. The first-order moment and second-order moment are corrected using the bias correction factor 0.9, and the corrected first-order moment estimate is 0.135 and the second-order moment estimate is 0.0207. The warm-up learning rate 7.744e-13 is divided by the square root of the corrected second-order moment estimate 0.144, and then multiplied by the corrected first-order moment estimate 0.135, plus the weight decay term 0.0001, and finally the momentum term is 7.27e-13.

[0180] System parameters are divided into three categories according to update requirements: fine-grained parameters include refrigerant flow control parameters, compressor speed parameters, etc., and the update frequency is once every 10 steps; medium-grained parameters include condenser temperature control parameters, evaporator pressure control parameters, etc., and the update frequency is once every 50 steps; coarse-grained parameters include system operation mode switching parameters, long-term optimization target parameters, etc., and the update frequency is once every 200 steps. The basic update frequency is set to 10 steps, and the policy gradients of fine-grained parameters, medium-grained parameters, and coarse-grained parameters are calculated to be 0.25, 0.15, and 0.1, respectively. These policy gradients are multiplied by weight coefficients 0.5, 0.3, and 0.2, respectively, and the weighted policy gradient is 0.185.

[0181] Gradient clipping is performed on the weighted policy gradient of 0.185, and the gradient norm threshold is set to 0.2. After clipping, the policy gradient value is 0.18. Based on the current system performance indicators, the learning rate weight 0.4, the momentum term weight 0.35, and the decay weight 0.25 are calculated. The product of the warm-up learning rate 7.744e-13 and the clipped policy gradient 0.18 is 1.39e-13, which is multiplied by the parameter update weight 0.4 to obtain 5.56e-14; the momentum term 7.27e-13 is multiplied by its parameter update weight 0.35 to obtain 2.54e-13; the original parameter value 0.5 is multiplied by its parameter update weight 0.25 to obtain 0.125. Sum these three items and finally obtain the optimized water production policy parameter value of 0.125000000000304.

[0182] Figure 3 : is a comparison diagram of the adaptive learning rate adjustment effect of the air water production intelligent prediction and regulation method based on meteorological big data in an embodiment of the present invention, such as Figure 3The figure shows the relationship between the learning rate adjustment characteristics and reward gain of different algorithms. The learning rate adjustment (diamond mark) of this technical solution shows a dynamic characteristic that is highly correlated with the reward gain (circle mark). In the early stage of training (0-200 steps), when the reward gain dropped from 2.5 to 1.8, the learning rate dropped steadily from 0.01 to 0.0085, showing the stabilizing effect of the warm-up mechanism. In the mid-term stage (300-600 steps), as the reward gain dropped from 1.5 to 0.85, the learning rate showed a more obvious downward trend, from 0.0073 to 0.0043. In the later stage of training (700-1000 steps), when the reward gain tends to stabilize in the range of 0.75-0.63, the adjustment of the learning rate is more precise, slowly decreasing from 0.0035 to 0.002. In contrast, the learning rate of the AdaBelief algorithm (square mark) is fixed at 0.0048 after 400 steps, and loses its adaptive ability; the RAdam algorithm (upper triangle mark) presents a simple linear attenuation feature, decreasing from 0.01 to 0.005, which is not strongly correlated with the improvement of system performance. This technical solution achieves a better adaptive optimization process by closely combining the adjustment of the learning rate with the change of the reward gain. In particular, in the interval where the reward gain changes dramatically (200-400 steps), the adjustment of the learning rate shows better follow-up and accuracy, which fully reflects the adaptive performance advantage of the algorithm.

[0183] In this embodiment, by introducing an adaptive step size and a warm-up mechanism, instability caused by excessive parameter updates in the early stage of training is avoided, the stability and convergence of the model training are improved, momentum estimation and bias correction mechanisms are adopted to overcome the problems of gradient sparsity and noise, parameter updates are smoother, and the model convergence speed is accelerated. In combination with multi-scale parameter optimization and gradient correction, hierarchical dynamic update of parameters is achieved, and the model's learning ability and generalization performance for features of different scales are improved;

[0184] The control strategy of the existing air-to-water system usually adopts a fixed learning rate and a unified parameter update frequency, and adopts the same update strategy for all control parameters. It is impossible to make differentiated adjustments to the response characteristics of different parameters, resulting in slow system response or over-adjustment. The fixed learning rate setting makes it difficult for the system to adaptively adjust the optimization step size during the training process, and it is easy to fall into the local optimal solution or produce training instability. In this embodiment, by constructing an adaptive step size adjustment function and a periodic learning rate scheduling function, dynamic adjustment of the learning rate is achieved, so that the system can automatically adjust the optimization step size according to real-time reward feedback. The introduction of a warm-up mechanism and a momentum term effectively suppresses the drastic fluctuations of parameters in the early stage of training and improves the stability of the training process. The multi-level parameter division method enables the system to respond to environmental changes on different time scales, which not only ensures the timely adjustment of fast parameters, but also maintains the stability of slow parameters. In summary, this embodiment achieves the unity of short-term control accuracy and long-term operation stability, greatly improves the air-to-water efficiency, reduces energy consumption, and is significantly better than the traditional single update strategy.

[0185] The intelligent prediction and regulation system of air-to-water production based on meteorological big data includes:

[0186] The first unit is used to collect meteorological monitoring data through meteorological sensors and perform standardization processing to generate a standardized meteorological feature matrix, extract periodic feature vectors and trend feature vectors through time series decomposition, and construct an enhanced feature space;

[0187] The second unit is used to build a transfer cross-domain learning network and add the pre-trained model knowledge in the fields of refrigeration and air conditioning and weather forecasting to the air-to-water prediction task through the knowledge transfer algorithm, and combine knowledge distillation to integrate expert knowledge from multiple source domains into the deep neural network;

[0188] The third unit is used to train a deep neural network based on the enhanced feature space and the back propagation algorithm and output the predicted water production volume hour by hour in the next day, use the predicted water production volume as the input of the state space and configure the action space, build a reward function according to the water production efficiency and energy consumption ratio, and build a deep reinforcement learning model based on the state space, action space and reward function;

[0189] The fourth unit is used to dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework and construct training samples corresponding to extreme meteorological conditions through an adversarial sample generator, optimize the deep reinforcement learning model based on the training samples to obtain a candidate water production parameter control strategy set, execute the water production process based on the candidate water production parameter control strategy set and set the optimization target;

[0190] The fifth unit is used to perform joint optimization through a multi-objective Bayesian optimizer, generate a Pareto optimal solution set, select optimal control parameters through a dynamic decision-making algorithm based on a confidence boundary, and dynamically adjust the update step size of the deep reinforcement learning model based on the optimal control parameters to obtain the optimal water production strategy.

[0191] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent prediction and regulation method for air-to-water production based on meteorological big data, characterized in that: include: Meteorological monitoring data is collected through meteorological sensors and standardized to generate a standardized meteorological feature matrix. Periodic feature vectors and trend feature vectors are extracted through time series decomposition and an enhanced feature space is constructed. Construct a transfer cross-domain learning network and add the pre-trained model knowledge in the fields of refrigeration and air conditioning and weather forecasting to the air-to-water prediction task through the knowledge transfer algorithm. Combine knowledge distillation to integrate expert knowledge from multiple source domains into the deep neural network. Based on the enhanced feature space, a deep neural network is trained in combination with a back propagation algorithm to output the predicted water production volume hour by hour in the next day, the predicted water production volume is used as the input of the state space and the action space is configured, a reward function is constructed according to the water production efficiency and energy consumption ratio, and a deep reinforcement learning model is constructed based on the state space, the action space and the reward function; Dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework and construct training samples corresponding to extreme meteorological conditions through an adversarial sample generator, optimize the deep reinforcement learning model based on the training samples to obtain a set of candidate water production parameter control strategies, execute the water production process based on the candidate water production parameter control strategy set and set the optimization target; A multi-objective Bayesian optimizer is used for joint optimization to generate a Pareto optimal solution set, and the optimal control parameters are selected through a dynamic decision-making algorithm based on a confidence boundary. The update step size of the deep reinforcement learning model is dynamically adjusted based on the optimal control parameters to obtain the optimal water production strategy.

2. The method according to claim 1, characterized in that Meteorological monitoring data is collected through meteorological sensors and standardized to generate a standardized meteorological feature matrix. Periodic feature vectors and trend feature vectors are extracted through time series decomposition and an enhanced feature space is constructed, including: Collect meteorological monitoring data through meteorological sensors, perform standardization on the meteorological monitoring data, and arrange the standardized meteorological monitoring data in time series to generate a standardized meteorological feature matrix; Performing time series decomposition on the standardized meteorological feature matrix, constructing a trajectory matrix and performing singular value decomposition, and decomposing the standardized meteorological feature matrix into a periodic component and a trend component, extracting a periodic feature vector and a trend feature vector of the standardized meteorological feature matrix; An enhanced feature space is constructed based on the periodic feature vector and the trend feature vector, wherein the enhanced feature space is used to characterize the periodic variation law and long-term variation trend of the meteorological monitoring data.

3. The method according to claim 1, characterized in that Construct a migration cross-domain learning network and add the pre-trained model knowledge in the refrigeration and air conditioning field and the weather forecast field to the air-to-water prediction task through the knowledge transfer algorithm. Combine knowledge distillation to integrate expert knowledge from multiple source domains into the deep neural network, including: Constructing a transfer cross-domain learning network, the transfer cross-domain learning network includes a feature extractor, a task-specific classifier and a domain discriminator, the feature extractor adopts a convolutional neural network structure with residual connection, the domain discriminator adopts a three-layer fully connected network structure, and the feature extractor is used to extract feature vectors from the refrigeration and air-conditioning field and the weather forecast field; Acquire feature statistical information based on the feature vector and input it into a kernel function generation network to obtain a kernel bandwidth parameter and a kernel function combination weight, identify the source domain of the feature vector based on a domain adversarial loss function to obtain a domain adversarial loss value, construct a maximum average difference metric based on the kernel bandwidth parameter and the kernel function combination weight, and calculate the difference between the feature vectors in the refrigeration and air conditioning field and the weather forecast field to obtain a maximum average difference value; For each feature vector, an attention mechanism is used to calculate the importance of the pre-trained model knowledge in the refrigeration and air-conditioning field and the weather forecast field to obtain a corresponding attention weight, and the attention weight is multiplied by the feature vector and superimposed to obtain a fusion feature; Based on the knowledge distillation loss function, the expert knowledge in the fusion feature is transferred to the deep neural network to obtain a knowledge distillation loss value, and the knowledge distillation loss value, the domain adversarial loss value and the maximum average difference value are weightedly combined to obtain an overall loss function; An alternating optimization strategy is adopted to train the transfer cross-domain learning network, the domain discriminator parameters corresponding to the domain discriminator are fixed and the feature extractor parameters corresponding to the feature extractor are optimized based on the overall loss function, the feature extractor parameters are fixed and the domain discriminator parameters are optimized based on the output of the domain discriminator, and the parameter optimization process is iteratively performed through the back propagation algorithm until the overall loss function converges to obtain a trained transfer cross-domain learning network.

4. The method according to claim 3, characterized in that Acquiring feature statistical information based on the feature vector and inputting it into a kernel function generation network to obtain a kernel bandwidth parameter and a kernel function combination weight, performing source domain identification on the feature vector based on a domain adversarial loss function to obtain a domain adversarial loss value, constructing a maximum average difference metric based on the kernel bandwidth parameter and the kernel function combination weight, and calculating the feature vector difference between the refrigeration and air conditioning field and the weather forecast field to obtain the maximum average difference value includes: Based on the feature vector, a feature pyramid structure is constructed in multiple network layers of a feature extractor, wherein the feature pyramid structure includes a shallow feature vector, a middle feature vector and a deep feature vector, wherein the shallow feature vector includes local texture information, the middle feature vector includes local semantic information, and the deep feature vector includes global semantic information; Calculating the mutual information between the shallow feature vector, the middle feature vector and the deep feature vector to obtain an inter-layer correlation score, and selecting the feature level combination with the strongest discriminability based on the inter-layer correlation score to obtain a combined feature vector; Acquire the mean and variance of the combined feature vector to construct feature statistical information, input the feature statistical information into a kernel function generation network, and the kernel function generation network outputs a kernel bandwidth parameter and a kernel function combination weight; Calculating the inter-class distance and intra-class distance of the combined feature vector in the refrigeration and air-conditioning field to obtain the feature distance in the refrigeration and air-conditioning field, and calculating the feature variance of the combined feature vector in the weather forecast field to obtain the feature variance in the weather forecast field; Calculating the feature dimension importance based on the feature distance in the refrigeration and air conditioning field and the feature variance in the weather forecast field, and weighting the feature dimension importance with the combined feature vector to obtain a weighted feature vector; The weighted feature vector is input into a domain discriminator, and the source domain of the weighted feature vector is identified based on a domain adversarial loss function to obtain a domain adversarial loss value. A maximum average difference metric is constructed based on the kernel bandwidth parameter and the kernel function combination weight. The maximum average difference metric is used to calculate the difference between the weighted feature vectors in the refrigeration and air-conditioning field and the meteorological forecast field to obtain a maximum average difference value.

5. The method according to claim 1, characterized in that Based on the enhanced feature space, the deep neural network is trained in combination with the back propagation algorithm and the predicted water production volume is output hour by hour in the next day. The predicted water production volume is used as the input of the state space and the action space is configured. The reward function is constructed according to the water production efficiency and energy consumption ratio. The deep reinforcement learning model is constructed based on the state space, action space and reward function, including: A deep neural network is constructed based on the enhanced feature space, wherein the deep neural network includes an input layer, three hidden layers and an output layer, wherein the hidden layer uses a rectified linear unit activation function; a back propagation algorithm is used to train the deep neural network to obtain hourly predicted water production in the next day; The predicted water production volume is constructed as the input of the state space, and an action space including compressor frequency, fan speed and system start / stop state is configured; a reward function is constructed based on water production efficiency and energy consumption ratio, wherein the water production efficiency is the water production volume per unit power, and the energy consumption ratio is the water production volume per unit energy consumption; A deep reinforcement learning model is constructed based on the state space, the action space and the reward function, wherein the deep reinforcement learning model includes an input layer, two hidden layers and an output layer; an experience replay buffer is set to store state-action transfer samples, and a greedy strategy is constructed for action exploration.

6. The method according to claim 1, characterized in that Dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework and construct training samples corresponding to extreme meteorological conditions through an adversarial sample generator, optimize the deep reinforcement learning model based on the training samples to obtain a candidate water production parameter control strategy set, execute the water production process based on the candidate water production parameter control strategy set and set the optimization target, including: Constructing a meta-learning adaptive framework, calculating a network structure adjustment gradient based on the adaptive framework, and dynamically adjusting the depth, number of neurons, and connection mode of the deep reinforcement learning model according to the network structure adjustment gradient to obtain an improved deep reinforcement learning model; Construct an adversarial sample generator, input the sample data under standard meteorological conditions into the adversarial sample generator, and generate training samples corresponding to extreme meteorological conditions of high temperature, low temperature, high humidity and low humidity through adversarial training; The improved deep reinforcement learning model is optimized and trained based on the training samples using an online learning method, and the model parameters are updated by minimizing the expected value of the reward function to obtain a set of candidate water production parameter control strategies under extreme meteorological conditions.

7. The method according to claim 1, characterized in that The multi-objective Bayesian optimizer is used for joint optimization to generate a Pareto optimal solution set and select the optimal control parameters through a dynamic decision-making algorithm based on a confidence boundary. The update step size of the deep reinforcement learning model is dynamically adjusted based on the optimal control parameters to obtain the optimal water production strategy, including: Construct a multi-objective Bayesian optimizer, take the water production efficiency objective function and the energy consumption ratio objective function as the optimization objectives of the multi-objective Bayesian optimizer, establish proxy models of the water production efficiency objective function and the energy consumption ratio objective function based on a Gaussian process regression model, use a radial basis function kernel to construct a kernel function of the proxy model, and perform sampling optimization on the proxy model based on an expected improvement criterion; A Pareto front is constructed based on a set of sampling points obtained by the multi-objective Bayesian optimizer, a non-dominated solution is used as a Pareto optimal solution set corresponding to the Pareto front, and a confidence boundary dynamic decision algorithm is constructed. The confidence boundary dynamic decision algorithm calculates a confidence upper bound and a confidence lower bound of each solution in the Pareto optimal solution set based on a confidence parameter, the confidence parameter is dynamically adjusted with the number of iterations, and a solution with the smallest uncertainty in the Pareto optimal solution set is selected as an optimal control parameter based on the difference between the confidence upper bound and the confidence lower bound; An adaptive step size adjustment function is designed based on the optimal control parameters, and the adaptive step size adjustment function is inversely proportional to the reward increment. The step size value and policy gradient obtained by the adaptive step size adjustment function are used to update the parameters of the deep reinforcement learning model. A momentum term is introduced to perform weighted averaging on the policy gradient. Based on the weighted policy gradient and step size, the parameters of the deep reinforcement learning model are updated in combination with the gradient correction enhancement mechanism and multi-scale parameter optimization to obtain the optimal water production strategy.

8. The method according to claim 7, characterized in that The step value and policy gradient obtained by the adaptive step adjustment function are used to update the parameters of the deep reinforcement learning model, a momentum term is introduced to perform weighted average on the policy gradient, and based on the weighted policy gradient and the step value, the parameters of the deep reinforcement learning model are updated in combination with the gradient correction enhancement mechanism and the multi-scale parameter optimization to obtain the optimal water production strategy, including: A step size value is obtained based on an adaptive step size adjustment function, a periodic learning rate scheduling function is constructed, a periodic learning rate value is calculated based on a basic learning rate and a maximum learning rate, the periodic learning rate value is multiplied by the step size value and the inverse of a reward gain to obtain a combined learning rate, and in the initial stage of training, a ratio of the combined learning rate to a preset time step is used as a warm-up coefficient, and the combined learning rate is multiplied by the warm-up coefficient to obtain a warm-up learning rate; Calculate the first-order moment estimate and the second-order moment estimate of the policy gradient, perform deviation correction on the first-order moment estimate and the second-order moment estimate based on a preset deviation correction coefficient to obtain a corrected first-order moment estimate and a corrected second-order moment estimate, multiply the warm-up learning rate by the ratio of the square roots of the corrected first-order moment estimate and the corrected second-order moment estimate, and add the ratio to the weight attenuation term to obtain a momentum term; Dividing the parameters of the deep reinforcement learning model into fine-grained parameters, medium-grained parameters and coarse-grained parameters, calculating the update frequency of the parameters of different granularities based on the basic update frequency, calculating the policy gradients of the parameters of different granularities, and multiplying the policy gradients of the parameters of different granularities with corresponding weight coefficients and summing them to obtain a weighted policy gradient; The weighted policy gradient is gradient clipped to obtain a clipped policy gradient, and the parameter update weight is calculated based on the learning rate weight, the weight of the momentum term and the decay weight. The product of the preheated learning rate and the clipped policy gradient, the momentum term and the original parameter are respectively multiplied by the corresponding parameter update weight and the sum is obtained to obtain the optimal water production strategy.

9. An intelligent prediction and regulation system for air-to-water production based on meteorological big data, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect meteorological monitoring data through meteorological sensors and perform standardization processing to generate a standardized meteorological feature matrix, extract periodic feature vectors and trend feature vectors through time series decomposition, and construct an enhanced feature space; The second unit is used to build a transfer cross-domain learning network and add the pre-trained model knowledge in the fields of refrigeration and air conditioning and weather forecasting to the air-to-water prediction task through the knowledge transfer algorithm, and combine knowledge distillation to integrate expert knowledge from multiple source domains into the deep neural network; The third unit is used to train a deep neural network based on the enhanced feature space and the back propagation algorithm and output the predicted water production volume hour by hour in the next day, use the predicted water production volume as the input of the state space and configure the action space, build a reward function according to the water production efficiency and energy consumption ratio, and build a deep reinforcement learning model based on the state space, action space and reward function; The fourth unit is used to dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework and construct training samples corresponding to extreme meteorological conditions through an adversarial sample generator, optimize the deep reinforcement learning model based on the training samples to obtain a candidate water production parameter control strategy set, execute the water production process based on the candidate water production parameter control strategy set and set the optimization target; The fifth unit is used to perform joint optimization through a multi-objective Bayesian optimizer, generate a Pareto optimal solution set, select optimal control parameters through a dynamic decision-making algorithm based on a confidence boundary, and dynamically adjust the update step size of the deep reinforcement learning model based on the optimal control parameters to obtain the optimal water production strategy.

Images