Intelligent Prediction and Regulation Method and System for Air Water Production Based on Meteorological Big Data

Abstract

The present invention provides an intelligent prediction and regulation method and system for air water production based on meteorological big data, which relates to the technical field of air water production, and includes: collecting data through meteorological sensors and constructing an enhanced feature space, using a transfer cross-domain learning network to fuse multi-domain knowledge, combining a deep neural network to predict the water production amount, generating a control strategy based on a deep reinforcement learning model, and selecting optimal control parameters through a multi-objective Bayesian optimization and dynamic decision-making algorithm, which can accurately predict the air water production amount and achieve intelligent regulation, improve the water production efficiency, reduce energy consumption, and has strong environmental adaptability and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of air water production, and particularly to an intelligent prediction and adjustment method and system for air water production amount based on meteorological big data. Background Art

[0002] Air water production technology is an innovative technology for extracting water from the air. By the principle of refrigeration condensation, water vapor in the air is converted into liquid water, which has important application value in fields such as water supply in water-scarce areas and emergency disaster relief. The efficiency of air water production is mainly affected by meteorological factors such as environmental temperature and relative humidity. Reasonably predicting and adjusting water production parameters is of great significance for improving water production efficiency;

[0003] Currently, air water production systems usually adopt a fixed parameter control scheme based on empirical rules, or use a simple feedback control strategy for parameter adjustment. With the development of meteorological monitoring technology and artificial intelligence technology, machine learning methods are tried to be applied to optimize water production parameters to improve system efficiency.

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

[0005] Therefore, there is an urgent need for a solution to solve the problems existing in the prior art. Summary of the Invention

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

[0007] In the first aspect of the embodiments of the present invention, an intelligent prediction and adjustment method for air water production amount based on meteorological big data is provided, including:

[0008] Collect meteorological monitoring data through meteorological sensors and perform standardized 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;

[0009] Construct a transfer cross-domain learning network and add the pre-trained model knowledge in the refrigeration and air-conditioning field and the meteorological forecasting field to the air water production prediction task through a knowledge transfer algorithm, and fuse the expert knowledge of multiple source domains into a deep neural network in combination with knowledge distillation;

[0010] Based on the enhanced feature space, train a deep neural network by combining with the backpropagation algorithm and output the predicted water production per hour within the next day. Use the predicted water production as the input of the state space and configure the action space. Construct a reward function according to the water production efficiency and energy consumption ratio. Build a deep reinforcement learning model based on the state space, action space and reward function;

[0011] Dynamically adjust the network structure of the deep reinforcement learning model through the 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 set of water production parameter control strategies. Execute the water production process based on the candidate set of water production parameter control strategies and set the optimization goal;

[0012] Perform joint optimization through a multi-objective Bayesian optimizer, generate a Pareto optimal solution set and select the optimal control parameters through a dynamic decision-making algorithm based on confidence bounds. 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.

[0013] In an alternative embodiment,

[0014] 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 including:

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

[0016] Perform time series decomposition on the standardized meteorological feature matrix. By constructing a trajectory matrix and performing singular value decomposition, decompose the standardized meteorological feature matrix into a periodic component and a trend component, and extract the periodic feature vector and trend feature vector of the standardized meteorological feature matrix;

[0017] Construct an enhanced feature space based on the periodic feature vector and the trend feature vector, where the enhanced feature space is used to characterize the periodic change law and long-term change trend of the meteorological monitoring data.

[0018] In an alternative embodiment,

[0019] Construct a transfer cross-domain learning network and add the pre-trained model knowledge in the refrigeration and air-conditioning field and the meteorological forecasting field to the air water production prediction task through the knowledge transfer algorithm. Combine knowledge distillation to fuse the expert knowledge of multiple source domains into the deep neural network including:

[0020] Construct a transfer cross - domain learning network, where 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 connections, and the domain discriminator adopts a three - layer fully - connected network structure. The feature extractor is used to extract feature vectors from the refrigeration and air - conditioning field and the weather forecasting field;

[0021] Obtain feature statistical information based on the feature vectors and input it into a kernel function generation network to get kernel bandwidth parameters and kernel function combination weights. Identify the source domain of the feature vectors based on the domain adversarial loss function to obtain a domain adversarial loss value. Construct a maximum mean discrepancy metric based on the kernel bandwidth parameters and the kernel function combination weights, and calculate the difference between the feature vectors of the refrigeration and air - conditioning field and the weather forecasting field to get a maximum mean discrepancy value;

[0022] For each feature vector, use an attention mechanism to calculate the importance degree of the pre - trained model knowledge in the refrigeration and air - conditioning field and the weather forecasting field to obtain corresponding attention weights, multiply the attention weights by the feature vectors and sum them up to obtain a fused feature;

[0023] Transfer the expert knowledge in the fused feature to a deep neural network based on the knowledge distillation loss function to obtain a knowledge distillation loss value, and perform weighted combination on the knowledge distillation loss value, the domain adversarial loss value, and the maximum mean discrepancy value to obtain an overall loss function;

[0024] Adopt an alternating optimization strategy to train the transfer cross - domain learning network. Fix the domain discriminator parameters corresponding to the domain discriminator and optimize the feature extractor parameters corresponding to the feature extractor based on the overall loss function. Fix the feature extractor parameters and optimize the domain discriminator parameters based on the output of the domain discriminator. Iteratively execute the parameter optimization process through the backpropagation algorithm until the overall loss function converges to obtain a trained transfer cross - domain learning network.

[0025] In an alternative embodiment,

[0026] Obtain feature statistical information based on the feature vectors and input it into a kernel function generation network to get kernel bandwidth parameters and kernel function combination weights. Identify the source domain of the feature vectors based on the domain adversarial loss function to obtain a domain adversarial loss value. Construct a maximum mean discrepancy metric based on the kernel bandwidth parameters and the kernel function combination weights, and calculate the difference between the feature vectors of the refrigeration and air - conditioning field and the weather forecasting field to get a maximum mean discrepancy value, including:

[0027] Construct a feature pyramid structure based on the feature vectors in multiple network layers of the feature extractor. The feature pyramid structure includes shallow feature vectors, middle feature vectors, and deep feature vectors. The shallow feature vectors contain local texture information, the middle feature vectors contain local semantic information, and the deep feature vectors contain global semantic information;

[0028] Calculate the mutual information between the shallow feature vectors, the middle feature vectors, and the deep feature vectors to obtain an inter-layer correlation score, and select the most discriminative feature level combination based on the inter-layer correlation score to obtain a combined feature vector;

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

[0030] Calculate the inter-class distance and intra-class distance of the combined feature vectors in the refrigeration and air-conditioning field to obtain the feature distance in the refrigeration and air-conditioning field, and calculate the feature variance of the combined feature vectors in the weather forecasting field to obtain the feature variance in the weather forecasting field;

[0031] Calculate the feature dimension importance based on the feature distance in the refrigeration and air-conditioning field and the feature variance in the weather forecasting field, and weight the feature dimension importance with the combined feature vector to obtain a weighted feature vector;

[0032] Input the weighted feature vector into a domain discriminator, identify the source domain of the weighted feature vector based on a domain adversarial loss function to obtain a domain adversarial loss value, construct a maximum mean discrepancy metric based on the kernel bandwidth parameter and the kernel function combination weight, and calculate the difference between the weighted feature vectors in the refrigeration and air-conditioning field and the weather forecasting field using the maximum mean discrepancy metric to obtain a maximum mean discrepancy value.

[0033] In an alternative embodiment,

[0034] Based on the enhanced feature space, train a deep neural network in combination with the backpropagation algorithm and output the predicted water production per hour within the next day. Use the predicted water production as the input of the state space and configure the action space, construct a reward function based on the water production efficiency and energy consumption ratio, and construct a deep reinforcement learning model based on the state space, action space, and reward function, including:

[0035] Construct a deep neural network based on the enhanced feature space. The deep neural network includes an input layer, three hidden layers, and an output layer. The hidden layer uses a rectified linear unit activation function; train the deep neural network using the backpropagation algorithm to obtain the predicted water production per hour within the next day;

[0036] Construct the predicted water production as the input of the state space, and configure an action space including compressor frequency, fan speed, and system start-stop status; construct a reward function based on water production efficiency and energy consumption ratio, where the water production efficiency is the water production per unit power, and the energy consumption ratio is the water production per unit energy consumption;

[0037] Construct a deep reinforcement learning model based on the state space, the action space, and the reward function. The deep reinforcement learning model includes an input layer, two hidden layers, and an output layer; set an experience replay buffer to store state-action transition samples, and construct a greedy policy for action exploration.

[0038] In an alternative implementation,

[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. Based on the candidate water production parameter control strategy set, execute the water production process and set the optimization objectives including:

[0040] Construct a meta-learning adaptive framework, calculate the network structure adjustment gradient based on the adaptive framework, and dynamically adjust the depth, number of neurons, and connection method 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 sampling data under standard meteorological conditions into the adversarial sample generator, and generate training samples corresponding to extreme meteorological conditions such as high temperature, low temperature, high humidity, and low humidity through an adversarial training method;

[0042] Adopt an online learning method to optimize and train the improved deep reinforcement learning model based on the training samples, and update the model parameters by minimizing the expected value of the reward function to obtain a candidate water production parameter control strategy set under extreme meteorological conditions.

[0043] In an alternative implementation,

[0044] Perform joint optimization through a multi-objective Bayesian optimizer, generate a Pareto optimal solution set, and select the optimal control parameters through a dynamic decision-making algorithm based on confidence bounds. 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 including:

[0045] Construct a multi-objective Bayesian optimizer, taking the water production efficiency objective function and the energy consumption ratio objective function as the optimization objectives of the multi-objective Bayesian optimizer. Based on the Gaussian process regression model, establish surrogate models for the water production efficiency objective function and the energy consumption ratio objective function. Use the radial basis function kernel to construct the kernel function of the surrogate model, and sample and optimize the surrogate model based on the expected improvement criterion;

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

[0047] Design an adaptive step size adjustment function based on the optimal control parameter. The adaptive step size adjustment function is inversely proportional to the reward increment. Use the step size value obtained by the adaptive step size adjustment function and the policy gradient to update the parameters of the deep reinforcement learning model. Introduce a momentum term to perform weighted averaging on the policy gradient. Based on the weighted policy gradient and the step size value, combine the gradient correction enhancement mechanism and multi-scale parameter optimization to update the parameters of the deep reinforcement learning model to obtain the optimal water production strategy.

[0048] In an alternative embodiment,

[0049] Using the step size value obtained by the adaptive step size adjustment function and the policy gradient to update the parameters of the deep reinforcement learning model, introducing a momentum term to perform weighted averaging on the policy gradient, and combining the gradient correction enhancement mechanism and multi-scale parameter optimization based on the weighted policy gradient and the step size value to update the parameters of the deep reinforcement learning model to obtain the optimal water production strategy includes:

[0050] Obtain the step size value based on the adaptive step size adjustment function, construct a periodic learning rate scheduling function, calculate the periodic learning rate value based on the base learning rate and the maximum learning rate, multiply the periodic learning rate value by the reciprocal of the step size value and the reward gain to obtain a combined learning rate, use the ratio of the combined learning rate to the preset time step as the warm-up coefficient at the initial stage of training, and multiply the combined learning rate by the warm-up coefficient to obtain the warm-up learning rate;

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

[0052] Divide the parameters of the deep reinforcement learning model into fine-grained parameters, medium-grained parameters, and coarse-grained parameters, calculate the update frequencies of different-grained parameters based on the basic update frequency, calculate the policy gradients of different-grained parameters, and multiply and sum the policy gradients of different-grained parameters with corresponding weight coefficients to obtain the weighted policy gradient;

[0053] Perform gradient clipping on the weighted policy gradient to obtain the clipped policy gradient, calculate the parameter update weight based on the learning rate weight, the weight of the momentum term, and the decay weight, and multiply and sum the product of the warm-up learning rate and the clipped policy gradient, the momentum term, and the original parameters with the corresponding parameter update weights to obtain the optimal water production strategy.

[0054] In a second aspect of the embodiments of the present invention, there is provided an intelligent prediction and regulation system for air water production based on meteorological big data, including:

[0055] A first unit, configured to collect meteorological monitoring data through meteorological sensors and perform standardized 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] A second unit, configured to construct a transfer cross-domain learning network and add the pre-trained model knowledge in the refrigeration and air-conditioning field and the meteorological forecasting field to the air water production prediction task through a knowledge transfer algorithm, and fuse the expert knowledge of multiple source domains into a deep neural network through knowledge distillation;

[0057] A third unit, configured to train a deep neural network based on the enhanced feature space in combination with the backpropagation algorithm and output the predicted water production per hour within the next day, use the predicted water production as the input of the state space and configure the action space, construct a reward function according to the water production efficiency and the energy consumption ratio, and construct a deep reinforcement learning model based on the state space, the action space, and the reward function;

[0058] A fourth unit, configured 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, and execute the water production process based on the candidate water production parameter control strategy set and set an optimization target;

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

[0060] In the present invention, by constructing a transfer cross-domain learning network, the pre-trained model knowledge in the fields of refrigeration and air-conditioning and meteorological forecasting is transferred to the air water production prediction task, and the expert knowledge of multiple source domains is fused by combining knowledge distillation technology, significantly improving the prediction accuracy and generalization ability of the model, effectively solving the problems of data sparsity and annotation difficulty in the air water production scenario, adopting a meta-learning adaptive framework to dynamically adjust the network structure of the deep reinforcement learning model, and constructing training samples corresponding to extreme meteorological conditions through an adversarial sample generator, enhancing the adaptability and robustness of the model in a complex and changeable environment, improving the control accuracy and stability of the water production process, and realizing the adaptive adjustment of water production parameters through joint optimization based on a multi-objective Bayesian optimizer and combining with a dynamic decision-making algorithm based on confidence bounds, reducing energy consumption while ensuring water production efficiency, improving the overall operation efficiency of the system, and having good engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic flowchart of the intelligent prediction and adjustment method for air water production volume based on meteorological big data according to an embodiment of the present invention;

[0062] Figure 2 It is a characteristic characterization quality assessment diagram of the intelligent prediction and adjustment method for air water production volume based on meteorological big data according to an embodiment of the present invention;

[0063] Figure 3 It is a comparison diagram of the adaptive learning rate adjustment effect of the intelligent prediction and adjustment method for air water production volume based on meteorological big data according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

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

[0066] Figure 1 This is a schematic flowchart of the intelligent prediction and adjustment method for air water production based on meteorological big data according to an embodiment of the present invention. As Figure 1 shown, the method includes:

[0067] 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;

[0068] Construct a transfer cross-domain learning network and add the knowledge of pre-trained models in the refrigeration and air-conditioning field and the meteorological forecasting field to the air water production prediction task through a knowledge transfer algorithm. Combine knowledge distillation to fuse the expert knowledge of multiple source domains into a deep neural network;

[0069] Based on the enhanced feature space, train a deep neural network in combination with the backpropagation algorithm and output the predicted water production per hour within the next day. Use the predicted water production as the input of the state space and configure the action space. Construct a reward function according to the water production efficiency and energy consumption ratio. Construct a deep reinforcement learning model based on the state space, action space, and 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 candidate set of water production parameter control strategies. Execute the water production process based on the candidate set of water production parameter control strategies and set optimization goals;

[0071] Perform joint optimization through a multi-objective Bayesian optimizer, generate a Pareto optimal solution set, and select the optimal control parameters through a dynamic decision algorithm based on confidence bounds. 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.

[0072] In an alternative embodiment,

[0073] 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, including:

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

[0075] Perform time series decomposition on the standardized meteorological feature matrix. By constructing a trajectory matrix and performing singular value decomposition, decompose the standardized meteorological feature matrix into a periodic component and a trend component, and extract the periodic feature vector and the trend feature vector of the standardized meteorological feature matrix;

[0076] Construct an enhanced feature space based on the periodic feature vector and the trend feature vector, where the enhanced feature space is used to characterize the periodic change law and the long-term change trend of the meteorological monitoring data.

[0077] Collect meteorological monitoring data through meteorological sensors. Meteorological sensors include temperature sensors, humidity sensors, barometric pressure sensors, wind speed sensors, etc. The collected meteorological monitoring data includes meteorological elements such as temperature, humidity, barometric pressure, and wind speed. Taking temperature data as an example, collect one year's temperature data with a sampling frequency of once per hour, and obtain 8760 temperature values.

[0078] Perform standardization processing on the collected meteorological monitoring data. The standardization processing includes: performing normalization processing on the original data, mapping the data to the 0-1 interval; removing outliers and correcting the data outside the normal range; filling in missing values, and using interpolation methods to fill in the missing data. Taking temperature data as an example, map the temperature values from -20°C to 40°C to the 0-1 interval, remove the obviously abnormal temperature values, and use linear interpolation methods to supplement the missing temperature values.

[0079] Arrange the standardized meteorological monitoring data in a time series to generate a standardized meteorological feature matrix. Taking temperature data as an example, arrange the 8760 standardized temperature values in chronological order into an 8760×1 matrix.

[0080] Perform time series decomposition on the standardized meteorological feature matrix. First, construct a trajectory matrix, select a sliding window length of 24 (corresponding to one day), and reconstruct the original time series into a 365×24 trajectory matrix. Perform singular value decomposition on the trajectory matrix to obtain eigenvalues and eigenvectors. According to the eigenvalue magnitudes, select the main components to reconstruct the time series to obtain the periodic component and the trend component.

[0081] Extract the periodic feature vector and the trend feature vector. The periodic feature vector reflects the periodic change law of the data, such as the daily and seasonal changes of temperature; the trend feature vector reflects the long-term change trend of the data, such as the interannual change trend of temperature.

[0082] Construct an enhanced feature space based on the periodic feature vector and the trend feature vector. Combine the periodic feature vector and the trend feature vector to construct a multi-dimensional feature space for characterizing the periodic change law and the long-term change trend of meteorological data.

[0083] In this embodiment, by performing standardization processing and time series decomposition on meteorological monitoring data, the periodic variation law and long-term variation trend in the data can be effectively extracted, improving the accuracy and reliability of meteorological data analysis. The method of trajectory matrix reconstruction and singular value decomposition is used for time series decomposition, which can effectively separate the periodic component and trend component in meteorological data, providing a reliable data basis for subsequent meteorological prediction and analysis. The constructed enhanced feature space comprehensively considers the periodic variation and trend variation of meteorological data, can more comprehensively characterize the variation characteristics of meteorological data, and provides richer feature information for meteorological prediction and analysis.

[0084] In an alternative embodiment,

[0085] Construct a transfer cross-domain learning network and add the pre-trained model knowledge in the refrigeration and air-conditioning field and the meteorological forecasting field to the air water production prediction task through a knowledge transfer algorithm. Combining knowledge distillation to fuse the expert knowledge of multiple source domains into a deep neural network includes:

[0086] Construct 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 connections, and the domain discriminator adopts a three-layer fully connected network structure. The feature extractor is used to extract feature vectors from the refrigeration and air-conditioning field and the meteorological forecasting field;

[0087] Based on the feature vectors, obtain feature statistical information and input it into a kernel function generation network to obtain kernel bandwidth parameters and kernel function combination weights. Based on a domain adversarial loss function, perform source domain identification on the feature vectors to obtain a domain adversarial loss value. Based on the kernel bandwidth parameters and the kernel function combination weights, construct a maximum mean discrepancy metric, and calculate the feature vector difference between the refrigeration and air-conditioning field and the meteorological forecasting field to obtain a maximum mean discrepancy value;

[0088] For each feature vector, use an attention mechanism to calculate the importance degree of the pre-trained model knowledge in the refrigeration and air-conditioning field and the meteorological forecasting field to obtain corresponding attention weights, multiply the attention weights by the feature vectors and stack them to obtain a fused feature;

[0089] Based on a knowledge distillation loss function, transfer the expert knowledge in the fused feature to a deep neural network to obtain a knowledge distillation loss value, and perform weighted combination of the knowledge distillation loss value, the domain adversarial loss value, and the maximum mean discrepancy value to obtain an overall loss function;

[0090] The transfer cross - domain learning network is trained using an alternating optimization strategy. 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. Then, the feature extractor parameters are fixed, and the domain discriminator parameters are optimized based on the output of the domain discriminator. The parameter optimization process is iteratively executed through the backpropagation algorithm until the overall loss function converges, and the trained transfer cross - domain learning network is obtained.

[0091] The structure of the feature extractor is designed in detail. The design of the convolutional layer adopts a multi - branch structure. The main branch contains conventional 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 realized through direct mapping or projection mapping to ensure that the output of the main branch is added to the bypass branch after matching the sizes. After the convolution operation, batch normalization is immediately performed to normalize the feature map in the channel dimension and introduce learnable scaling and translation parameters.

[0092] For the design of the fully - connected layer of the domain discriminator, the number of neurons in the input layer corresponds to the output dimension of the feature extractor. The middle layer realizes feature compression by gradually reducing the dimension, and the output layer generates two - dimensional discrimination results. The weight matrix of the fully - connected layer is initialized using a uniform distribution, and the bias term is initialized to zero. A non - linear activation function is introduced after each linear transformation, and the rectified linear unit function is selected to introduce non - linear 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), higher - order moments (skewness, kurtosis), etc. Based on these statistical information, a kernel function parameter generator is constructed, and the generator outputs the kernel bandwidth parameter and the combination weight through a feed - forward 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 the domain adversarial loss is divided into two steps: the feature vector is input into the domain discriminator to obtain a two - dimensional output, and after normalization, the domain discrimination probability is obtained. The cross - entropy is calculated between the probability value and the true domain label. The calculation of the maximum mean discrepancy is based on the generated kernel function parameters and is realized by calculating the distance between the two domains in the reproducing kernel Hilbert space.

[0095] The implementation of the attention mechanism adopts the scaled dot - product attention structure. The feature vectors are respectively converted into query vectors, key vectors, and value vectors. The similarity scores are calculated through the dot - product of the query vectors and the key vectors, and the attention weights are obtained after scaling and normalization. The attention weights are multiplied by the value vectors to obtain the weighted features. The multi - head attention is realized by calculating multiple groups of attention results in parallel and concatenating 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 computing the relative entropy between the soft labels and the output of the student network. The overall loss function combines three loss terms through adjustable weight coefficients, and the weight coefficients can be adjusted according to the actual task requirements.

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

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

[0099] Exemplarily, in the field of refrigeration and air - conditioning systems of the air - to - water system, the original data includes the operating data of the refrigeration system, specifically including: the inlet and outlet temperatures, pressures, and flows of the compressor, the inlet and outlet temperatures, pressures, and flows of the condenser, the inlet and outlet temperatures, pressures, and flows of the evaporator, the opening degree of the expansion valve, and other operating parameters. In the field of weather forecasting, the original data includes meteorological element data such as ambient temperature, relative humidity, atmospheric pressure, wind speed, and dew - point temperature.

[0100] First, the data in the field of refrigeration and air - conditioning is input into the feature extractor. Through multi - layer convolution processing, the operating features of the refrigeration system are extracted, and these features reflect the working relationships and energy conversion laws among the components of the refrigeration system. Similarly, the data in the field of weather forecasting is input into the same feature extractor to extract the change features among meteorological elements, and these features reflect the dynamic change laws of the atmospheric environment.

[0101] The extracted feature vectors are input into the kernel function generation network, which 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 associated features between the compressor and other components; when the ambient temperature data is input, the network will pay more attention to the coupling features between temperature and humidity.

[0102] The domain discriminator identifies the source of the feature vectors, determining whether the features come from the refrigeration and air-conditioning field or the weather forecasting field. The difference degree between the features of the two fields is calculated through the maximum mean discrepancy, which is used to evaluate the domain adaptation ability of the feature extractor.

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

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

[0105] In the alternating optimization process, first optimize the feature extractor so that it can extract the common features of the refrigeration system and the meteorological environment, and then optimize the domain discriminator to improve its ability to distinguish features in different domains. For example, the feature extractor gradually learns the correlation law between the compressor power and the ambient temperature, and the domain discriminator learns to judge which domain this correlation law belongs to.

[0106] The finally trained network can comprehensively utilize the knowledge in the refrigeration and air-conditioning field and the weather forecasting field to achieve accurate prediction of water production. For example, when new environmental conditions and refrigeration system parameters are input, the network can predict the optimal water production working conditions and the corresponding water production. When the environmental conditions change, the network can adaptively adjust the control strategy to maintain a high water production efficiency.

[0107] In this embodiment, through transfer learning and knowledge distillation techniques, the pre-trained model knowledge in the refrigeration and air-conditioning field and the weather forecasting field is effectively utilized, improving the performance of the air water production prediction task. By adopting the attention mechanism and the kernel function adaptive selection strategy, the dynamic fusion of different source domain knowledge is realized, enhancing the adaptability of the model to different working conditions. Based on domain adversarial training and maximum mean discrepancy measurement, 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 decreased;

[0108] In the existing air water production system, the prediction of water production usually adopts a knowledge modeling method in a single field. Only based on the knowledge in the refrigeration and air conditioning field, a prediction model is constructed, taking the operating parameters of each component of the refrigeration system as inputs for modeling, ignoring the impact of meteorological environment changes on water production efficiency, or only relying on the knowledge in the meteorological forecast field, predicting the theoretical water production based on environmental parameters, without fully considering the actual operating characteristics of the refrigeration system. The modeling method in a single field is difficult to accurately describe the complex coupling relationship between the refrigeration system and the environment in the air water production process, resulting in insufficient prediction accuracy. Especially in extreme weather or under non-rated operating conditions of the equipment, the prediction error is relatively large;

[0109] In this embodiment, a convolutional neural network with residual connections is used to extract the deep features in the refrigeration and air conditioning field and the meteorological forecast field. The design of residual connections effectively alleviates the problem of gradient disappearance in deep networks and improves the accuracy of feature extraction. The introduction of the domain discriminator and the maximum mean discrepancy metric realizes the alignment of the feature distributions in different fields, making the extracted features have better domain adaptability. The knowledge distillation technology efficiently transfers the 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, providing strong support for the intelligent control of the air water production system.

[0110] In an alternative embodiment,

[0111] Based on the feature vector, obtain the feature statistical information and input it into the kernel function generation network to obtain the kernel bandwidth parameter and the kernel function combination weight. Based on the domain adversarial loss function, perform source domain identification on the feature vector to obtain the domain adversarial loss value. Based on the kernel bandwidth parameter and the kernel function combination weight, construct the maximum mean discrepancy metric, and calculate the feature vector difference between the refrigeration and air conditioning field and the meteorological forecast field to obtain the maximum mean discrepancy value, including:

[0112] Based on the feature vector, construct a feature pyramid structure in multiple network layers of the feature extractor. The feature pyramid structure includes shallow feature vectors, middle feature vectors, and deep feature vectors. The shallow feature vectors contain local texture information, the middle feature vectors contain local semantic information, and the deep feature vectors contain global semantic information;

[0113] Calculate the mutual information between the shallow feature vectors, the middle feature vectors, and the deep feature vectors to obtain the inter-layer correlation score. Based on the inter-layer correlation score, select the most discriminative feature level combination to obtain the combined feature vector;

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

[0115] Calculate the between-class distance and within-class distance of the combined feature vector in the field of refrigeration and air conditioning to obtain the feature distance in the field of refrigeration and air conditioning, and calculate the feature variance of the combined feature vector in the field of weather forecasting to obtain the feature variance in the field of weather forecasting;

[0116] Calculate the importance of feature dimensions based on the feature distance in the field of refrigeration and air conditioning and the feature variance in the field of weather forecasting, and weight the importance of feature dimensions with the combined feature vector to obtain a weighted feature vector;

[0117] Input the weighted feature vector into a domain discriminator, identify the source domain of the weighted feature vector based on the domain adversarial loss function to obtain a domain adversarial loss value, construct a maximum mean discrepancy metric based on the kernel bandwidth parameter and the kernel function combination weight, and calculate the difference between the weighted feature vectors in the field of refrigeration and air conditioning and the field of weather forecasting using the maximum mean discrepancy metric to obtain a maximum mean discrepancy value.

[0118] Construct a feature pyramid structure in different network layers of the feature extractor. In the shallow network, extract local texture information through convolutional operations with a small receptive field, capture low-level features such as edges and textures in the data, and form a shallow feature vector. In the middle layer of the network, as the receptive field expands, extract semantic information of local regions to obtain a more representative middle-layer feature vector. In the deep network, expand the receptive field range through multiple convolutional and downsampling operations to extract a deep feature vector with global semantic information.

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

[0120] For the combined feature vector, calculate its distribution characteristics in the feature space. First, obtain the mean and variance of the feature vector in each dimension to construct feature statistical information. Input this statistical information into a pre-designed kernel function generation network, which adaptively generates a kernel bandwidth parameter and a kernel function combination weight through multiple transformations.

[0121] In the field of refrigeration and air conditioning, calculate the inter-class distance and intra-class distance of the combined feature vectors. The inter-class distance reflects the discrimination of the features of different category samples, and the intra-class distance reflects the compactness of the features of the same category samples. Together, they constitute the feature distance in the field of refrigeration and air conditioning. For the field of weather forecasting, calculate the variance distribution of the combined feature vectors in each dimension to obtain the feature variance in the field of weather forecasting.

[0122] Based on the calculated feature distance in the field of refrigeration and air conditioning and the feature variance in the field of weather forecasting, evaluate the importance of each dimension in the feature space. Assign a higher importance score to the feature dimensions with strong discrimination and large variance. Perform a weighting operation on the obtained feature dimension importance and the combined feature vectors to obtain weighted feature vectors.

[0123] Input the weighted feature vectors into the domain discriminator for source domain identification. Calculate the domain discrimination loss value of the feature vectors through the domain adversarial loss function, and at the same time construct the maximum mean discrepancy metric based on the parameters output by the kernel function generation network. Use this metric to calculate the distribution difference between the weighted feature vectors in the field of refrigeration and air conditioning and the field of weather forecasting to obtain the maximum mean discrepancy value.

[0124] Exemplarily, in an 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 through multiple layers of the feature extractor:

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

[0126] By calculating the mutual information between the feature levels, it is found that the shallow features can better represent the instantaneous changes of the parameters, the middle features reflect the mutual influences between the parameters, and the deep features reflect the overall operating state of the system. Based on the complementarity of these features, select the most discriminative level combination to construct the combined feature vectors.

[0127] For the field of refrigeration and air conditioning, calculate the feature distances under different operating conditions, and the features of the high-efficiency and low-efficiency operating conditions show large inter-class differences. In the field of weather forecasting, the variance distribution of the features under different weather conditions reflects the change rules of the environmental parameters.

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

[0129] The weighted feature vector is input into the domain discriminator for recognition, and at the same time, the difference degree of the features in the two domains is evaluated through the maximum mean discrepancy metric, providing a basis for subsequent knowledge transfer and fusion.

[0130] Figure 2 This is the feature characterization quality evaluation diagram of the intelligent prediction and regulation method for air water production based on meteorological big data according to the embodiments of the present invention. As Figure 2 shown, the feature characterization quality evaluation diagram shows the comparison of the present technical solution with PCA and AutoEncoder in terms of the evaluation of the importance of feature dimensions. 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 and 0.88, significantly higher than that of PCA (0.62 - 0.68) and AutoEncoder (0.55 - 0.61). Especially at dimension 20, the present technical solution reaches the highest score of 0.87, while PCA and AutoEncoder are 0.67 and 0.59 respectively. In the secondary feature interval of dimension 30 - 40, the present technical solution shows a more delicate discrimination ability, and the importance score shows a reasonable fluctuation range (0.45 - 0.65), while the scores of other methods in this interval are relatively flat (PCA: 0.52 - 0.58, AutoEncoder: 0.42 - 0.48). Therefore, the present technical solution has significant advantages compared with 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 generation network is used to adaptively learn the kernel function parameters, avoiding the limitations of manual parameter setting, making the measurement method more flexible and accurate. Combining domain adversarial learning and maximum mean discrepancy metric not only ensures the discriminability of features but also reduces the domain difference, achieving a better cross-domain knowledge transfer effect;

[0132] The existing methods for extracting features of air water-making systems usually adopt a single-level feature extraction strategy, which cannot capture the multi-scale feature information of data simultaneously, resulting in limited feature expression ability. In the feature selection process, a fixed weight method is often used, making it difficult to dynamically adjust the importance of features according to the actual working conditions. At the same time, there is a lack of an effective domain adaptation mechanism when dealing with data from different fields, leading to a large domain shift in the extracted features and affecting the generalization ability of the model. In this embodiment, a feature pyramid structure is constructed in different network layers of the feature extractor to achieve multi-level feature extraction from local texture to global semantics. By calculating the mutual information between layers, the complementarity of features at different levels is evaluated, and the most discriminative feature combination is selected, effectively improving the feature expression ability. 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 weather forecasting, the adaptive calculation of feature importance is realized. In summary, this embodiment realizes the efficient feature extraction and domain adaptation of data in the fields of refrigeration and air conditioning and weather forecasting, significantly improving the feature expression ability, discriminative ability, and generalization ability, providing more reliable feature support for the performance prediction and optimal control of the air water-making system, and being of great significance for improving the overall performance of the system.

[0133] In an alternative embodiment,

[0134] Based on the enhanced feature space, train a deep neural network in combination with the backpropagation algorithm and output the predicted water production per hour within the next day. Use the predicted water production as the input of the state space and configure the action space. Construct a reward function based on the water production efficiency and energy consumption ratio. The deep reinforcement learning model constructed based on the state space, action space, and reward function includes:

[0135] Construct a deep neural network based on the enhanced feature space. The deep neural network includes an input layer, three hidden layers, and an output layer. The hidden layer uses a rectified linear unit activation function; train the deep neural network using the backpropagation algorithm to obtain the predicted water production per hour within the next day;

[0136] Construct the predicted water production as the input of the state space and configure an action space including the compressor frequency, fan speed, and system start-stop state; construct a reward function based on the water production efficiency and energy consumption ratio, where the water production efficiency is the water production per unit power, and the energy consumption ratio is the water production per unit energy consumption;

[0137] Construct a deep reinforcement learning model based on the state space, the action space, and the reward function. The deep reinforcement learning model includes an input layer, two hidden layers, and an output layer; set an experience replay buffer to store state-action transition samples, and construct a greedy policy for action exploration.

[0138] Build a deep neural network for predicting 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 the rectified linear unit is used as the activation function for all of them. The number of neurons in the output layer is 24, corresponding to the predicted water production for the next 24 hours. The network is trained using the backpropagation algorithm with a learning rate of 0.001 and 1000 training epochs. In each epoch, 80% of the samples are randomly selected for training.

[0139] Take the predicted 24-hour water production as the input to the state space, and configure the action space to include the compressor frequency (range 20 - 60 Hz, step size 5 Hz), the fan speed (range 500 - 1500 rpm, step size 100 rpm), and the system start / stop state (0 for stop, 1 for running). Define the water production efficiency based on the water production per unit power, with the unit of kg / kW; define the energy consumption ratio based on the water production per unit energy consumption, with the unit of kg / kWh. Construct a reward function by weighted summing the water production efficiency and the energy consumption ratio, with weights of 0.6 and 0.4 respectively.

[0140] Build a deep reinforcement learning model, which consists of an input layer, two hidden layers (with 64 and 32 neurons respectively), and an output layer. Set the size of the experience replay buffer to 10000, which is used to store the transition samples of states, actions, rewards, and next states. Use the ε-greedy strategy for action exploration, with the initial ε value of 0.9, decaying to 0.01 at a decay rate of 0.995. In each training epoch, 256 samples are randomly sampled for update, and the number of training epochs is 5000.

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

[0142] In this embodiment, an accurate prediction of future water production is achieved through a deep neural network, providing a reliable decision-making basis for system control. Combining deep reinforcement learning, intelligent optimization of system control parameters is realized. Using experience replay and the greedy strategy ensures the stability and exploration efficiency of model training, enabling the model to have good generalization ability and adaptive ability, and continuously optimizing the system operation effect.

[0143] In an alternative implementation,

[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, and execute the water production process based on the candidate water production parameter control strategy set. The set optimization objectives include:

[0145] Construct a meta-learning adaptive framework, calculate the network structure adjustment gradient based on the adaptive framework, and dynamically adjust the depth, number of neurons, and connection method 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 sampling data under standard meteorological conditions into the adversarial sample generator, and generate training samples corresponding to extreme meteorological conditions such as high temperature, low temperature, high humidity, and low humidity through an adversarial training method;

[0147] Adopt an online learning method to optimize and train the improved deep reinforcement learning model based on the training samples, and update the model parameters by minimizing the expected value of the reward function to obtain a candidate water production parameter control strategy set 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 structure adjustment by calculating the gradient of the loss on the validation set with respect to the network structure parameters. Specifically, when implementing, set the depth range to 3 - 10 layers, the number of neurons per layer range to 32 - 256, and the connection methods include residual connection and dense connection. According to the gradient update direction, gradually adjust the network depth, number of neurons, and connection method, and complete the structure optimization when the performance of the validation set no longer improves.

[0149] Construct an adversarial sample generator, which includes two networks: a generator and a discriminator. The generator adopts an encoder-decoder structure, and the input is data such as temperature and humidity under standard meteorological conditions, and generates extreme meteorological condition samples through a multi-layer convolutional and transposed convolutional network. The discriminator is used to judge the authenticity of the generated samples. During training, collect the meteorological data for 7 consecutive days under standard conditions as the training set, including 5000 data samples with a temperature of 20 - 25°C and a relative humidity of 50 - 70%. Generate 1000 training samples for each of the extreme conditions such as high temperature (35 - 40°C), low temperature (0 - 5°C), high humidity (85 - 95%), and low humidity (20 - 30%) through an adversarial training method.

[0150] Optimize the deep reinforcement learning model using online learning. Set the state space to include 12 parameters such as influent water quality and meteorological conditions, and the action space to include 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 training samples under extreme operating conditions is obtained, the model parameters are updated immediately. After 1000 rounds of iterative training, the average reward value of the model under various extreme operating conditions increases by more than 30%, and the corresponding optimal control strategy set is obtained.

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

[0152] In an alternative embodiment,

[0153] Joint optimization is performed through a multi-objective Bayesian optimizer to generate a Pareto optimal solution set and select the optimal control parameters through a dynamic decision-making algorithm based on confidence bounds. Dynamically adjusting the update step size of the deep reinforcement learning model based on the optimal control parameters to obtain the optimal water production strategy includes:

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

[0155] Construct a Pareto front based on the set of sampling points obtained by the multi-objective Bayesian optimizer, take the non-dominated solutions as the Pareto optimal solution set corresponding to the Pareto front, construct a confidence bound dynamic decision-making algorithm, the confidence bound dynamic decision-making algorithm calculates the upper confidence bound and the lower confidence bound of each solution in the Pareto optimal solution set based on the confidence parameter, the confidence parameter is dynamically adjusted with the number of iterations, and select the solution with the least uncertainty in the Pareto optimal solution set as the optimal control parameter based on the difference between the upper confidence bound and the lower confidence bound;

[0156] Design an adaptive step size adjustment function based on the optimal control parameters. The adaptive step size adjustment function is inversely proportional to the reward increment. Use the step size value obtained from the adaptive step size adjustment function and the policy gradient to update the parameters of the deep reinforcement learning model. Introduce a momentum term to perform weighted averaging on the policy gradient. Based on the weighted policy gradient and the step size value, combine the gradient correction enhancement mechanism and multi-scale parameter optimization to update the parameters of the deep reinforcement learning model to obtain the optimal water production strategy.

[0157] Construct a multi-objective Bayesian optimizer. Take water production efficiency and energy consumption ratio as two optimization objectives, and establish their corresponding surrogate models respectively. The surrogate model is implemented based on Gaussian process regression, and the radial basis function kernel is used to describe the correlation between sample points. For each surrogate model, determine the next most valuable sampling point based on the expected improvement criterion, which takes into account both the predicted mean and variance to balance exploration and exploitation.

[0158] After obtaining the set of sampling points, construct the Pareto front. Perform non-dominated sorting on all sampling points, and take the solutions that are not dominated by any other solutions as the Pareto optimal solution set. Construct a confidence bound dynamic decision-making algorithm, which introduces a confidence parameter to balance the stability and adaptability of the decision-making. For each solution in the Pareto optimal solution set, calculate its upper and lower confidence bounds based on the confidence parameter. The confidence parameter will be dynamically adjusted during the iteration process, being larger in the initial stage of optimization to encourage exploration and gradually decreasing in the later stage to promote convergence. Evaluate the uncertainty of each solution by calculating the difference of the confidence bounds, and select the solution with the minimum uncertainty as the optimal control parameter.

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

[0160] During the parameter update process, combine the gradient correction enhancement mechanism, normalize the gradient and introduce an adaptive learning rate. Through the multi-scale parameter optimization strategy, optimize the model parameters simultaneously at different scales to improve the optimization efficiency. Finally, obtain the optimal water production strategy that can balance water production efficiency and energy consumption.

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

[0162] A surrogate model is established through Gaussian process regression to predict the water production efficiency and energy consumption ratio under different combinations of control parameters. 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 predicted water production efficiency at a certain operating condition point is high and the prediction uncertainty is large, this point will be selected as the next sampling point.

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

[0164] For the selected optimal control parameters, an adaptive step size adjustment function is designed. When the system achieves a significant performance improvement under a certain operating 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 new control strategies. After introducing the momentum term, the update of the control parameters will consider the historical gradient information to avoid falling into local optima.

[0165] Through gradient correction and multi-scale optimization, the control strategy is optimized at different time scales. For example, the fast response of the compressor frequency is optimized at a short time scale, and the switching strategy of the system operating mode is optimized at a long time scale. The finally obtained water production strategy can adaptively adjust the control parameters according to the environmental conditions and system states, achieving the dynamic balance of water production efficiency and energy consumption.

[0166] In this embodiment, the method combining multi-objective Bayesian optimization and Gaussian process regression can efficiently construct a surrogate model of the objective function, significantly reduce the number of samplings, accelerate the optimization convergence speed, and through the confidence bound-based dynamic decision algorithm, by adaptively adjusting the confidence parameter, achieve a good balance between exploration and exploitation, effectively avoid falling into local optimal solutions, introduce the adaptive step size adjustment and momentum term mechanisms, and combine gradient correction and multi-scale optimization, significantly improving the training stability and convergence performance of the deep reinforcement learning model and ensuring the acquisition of the optimal water production strategy;

[0167] The control optimization methods of existing air water production systems are difficult to simultaneously consider the two objectives of water production efficiency and energy consumption ratio. Often, while improving the water production efficiency, the energy consumption increases significantly. Conventional multi-objective optimization algorithms are prone to falling into local optima when dealing with high-dimensional non-linear optimization problems, and the sampling efficiency is low. Existing reinforcement learning methods use a fixed step size in the parameter update process, making it difficult to adapt to the dynamic characteristics of the system. At the same time, there is a lack of effective guarantee for the stability of gradient update. In this embodiment, by constructing a multi-objective Bayesian optimizer, taking both the water production efficiency and the energy consumption ratio as optimization objectives, using a Gaussian process regression model to establish a surrogate model of the objective function, effectively capturing the non-linear relationship in the parameter space through a radial basis function kernel, and using the expected improvement criterion to achieve efficient sampling optimization. By constructing a Pareto front and retaining all non-dominated solutions, the subjectivity of artificially weighting multiple objectives in traditional methods is avoided. In summary, through the innovative combination of multi-objective Bayesian optimization, dynamic decision-making mechanism and adaptive deep reinforcement learning in this embodiment, the comprehensive optimization of the air water production system control strategy is realized, and significant improvements have been achieved in aspects such as water production efficiency and energy consumption balance, decision-making reliability, training stability and optimization efficiency, providing a more efficient and reliable solution for the intelligent control of air water production systems.

[0168] In an alternative embodiment,

[0169] Using the step size value obtained by the adaptive step size adjustment function and the policy gradient to update the parameters of the deep reinforcement learning model, introducing a momentum term to perform weighted averaging on the policy gradient, and based on the weighted policy gradient and the step size value, combining a gradient correction enhancement mechanism and multi-scale parameter optimization to update the parameters of the deep reinforcement learning model to obtain an optimal water production strategy includes:

[0170] Based on the adaptive step size adjustment function, obtain a step size value, construct a periodic learning rate scheduling function, calculate the periodic learning rate value based on the base learning rate and the maximum learning rate, multiply the periodic learning rate value by the reciprocal of the step size value and the reward gain to obtain a combined learning rate, use the ratio of the combined learning rate to a preset time step as a warm-up coefficient at the initial stage of training, and multiply the combined learning rate by the warm-up coefficient to obtain a warm-up learning rate;

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

[0172] Divide the parameters of the deep reinforcement learning model into fine-grained parameters, medium-grained parameters, and coarse-grained parameters. Calculate the update frequencies of different-grained parameters based on the base update frequency, calculate the policy gradients of different-grained parameters, multiply the policy gradients of different-grained parameters by the corresponding weight coefficients and sum them to obtain the weighted policy gradient.

[0173] Perform gradient clipping on the weighted policy gradient to obtain the clipped policy gradient. Calculate the parameter update weight based on the learning rate weight, the weight of the momentum term, and the decay weight. Multiply the product of the warm-up learning rate and the clipped policy gradient, the momentum term, and the original parameters by the corresponding parameter update weights respectively and sum them to obtain the optimal water production strategy.

[0174] The adaptive step size adjustment function calculates the step size value according to the system operating state and performance metrics. The periodic learning rate scheduling function calculates the periodic learning rate value based on the set base learning rate and maximum learning rate parameters according to the variation law of the training period. Multiply the calculated periodic learning rate value, the step size value, and the reciprocal of the current reward gain of the system to obtain a combined learning rate that can be adaptively adjusted. In the initial stage of training the deep reinforcement learning model, to ensure training stability, perform a ratio operation on the combined learning rate and the preset time step to obtain the warm-up coefficient, and multiply the combined learning rate by this warm-up coefficient to obtain the warm-up learning rate, realizing a smooth adjustment of the learning rate.

[0175] Perform moment estimation on the policy gradient of the deep reinforcement learning model to calculate the first-order moment estimation value and the second-order moment estimation value. To eliminate the cumulative bias in the estimation process, introduce a preset bias correction coefficient, and perform bias correction processing on the first-order moment estimation value and the second-order moment estimation value respectively to obtain the corrected first-order moment estimation value and the corrected second-order moment estimation value. Multiply the warm-up learning rate by the corrected first-order moment estimation value, then divide it by the square root of the corrected second-order moment estimation value, and perform an addition operation with 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 rapid response and fine adjustment, medium-grained parameters handle medium-time-scale changes, and coarse-grained parameters control long-term policy adjustments. Based on the set base update frequency, calculate the update frequencies of the three types of parameters respectively, and the update frequency decreases as the parameter granularity increases. Calculate the policy gradients of the three types of parameters respectively, and the policy gradient reflects the degree of influence of the parameter on the system performance. Multiply the policy gradients of different-grained parameters by their corresponding weight coefficients and sum them to obtain the weighted policy gradient that comprehensively considers multiple time scales.

[0177] Perform gradient clipping on the weighted policy gradient. By setting a gradient norm threshold, limit overly large gradient values within a reasonable range to obtain the clipped policy gradient, thus avoiding the problem of gradient explosion during the parameter update process. Based on the system performance metrics and training progress, calculate the learning rate weight, momentum term weight, and decay weight. These weights are used to balance the contributions of different update terms to obtain the parameter update weights. Multiply the product term of the warm-up learning rate and the clipped policy gradient, the momentum term containing historical information, and the original parameters reflecting parameter inertia by their corresponding parameter update weights respectively, and perform a summation operation on the obtained results to get the final optimal water production strategy.

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

[0179] The first moment estimate value of the policy gradient is calculated to be 0.15, and the second moment estimate value is 0.023. Use the bias correction coefficient of 0.9 to correct the first moment and the second moment, and the corrected first moment estimate value is 0.135, and the second moment estimate value is 0.0207. Divide the warm-up learning rate of 7.744e-13 by the square root of the corrected second moment estimate value of 0.144, then multiply by the corrected first moment estimate value of 0.135, and add the weight decay term of 0.0001. Finally, the momentum term is 7.27e-13.

[0180] System parameters are divided into three categories according to the 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. Based on this, the policy gradient of fine-grained parameters is calculated to be 0.25, the policy gradient of medium-grained parameters is 0.15, and the policy gradient of coarse-grained parameters is 0.1. Multiply these policy gradients by the weight coefficients 0.5, 0.3, and 0.2 respectively and sum them up to obtain the weighted policy gradient of 0.185.

[0181] Perform gradient clipping on the weighted policy gradient of 0.185, set the gradient norm threshold to 0.2, and obtain the policy gradient value of 0.18 after clipping. Based on the current system performance indicators, calculate the learning rate weight of 0.4, the momentum term weight of 0.35, and the decay weight of 0.25. The product of the warm-up learning rate of 7.744e-13 and the clipped policy gradient of 0.18 is 1.39e-13, and this value is multiplied by the parameter update weight of 0.4 to obtain 5.56e-14; the momentum term of 7.27e-13 is multiplied by its parameter update weight of 0.35 to obtain 2.54e-13; the original parameter value of 0.5 is multiplied by its parameter update weight of 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 This is a comparison chart of the adaptive learning rate adjustment effect of the air water production intelligent prediction and adjustment method based on meteorological big data in the embodiment of the present invention, as Figure 3The figure shows the relationship between the learning rate adjustment characteristics of different algorithms and the reward gain. The learning rate adjustment (diamond markers) of this technical solution exhibits dynamic characteristics highly correlated with the reward gain (circular markers). In the initial stage of training (0 - 200 steps), when the reward gain decreases from 2.5 to 1.8, the learning rate steadily drops from 0.01 to 0.0085, demonstrating the stable effect of the warm-up mechanism. In the middle stage (300 - 600 steps), as the reward gain decreases from 1.5 to 0.85, the learning rate shows a more significant downward trend, dropping from 0.0073 to 0.0043. In the later stage of training (700 - 1000 steps), when the reward gain stabilizes in the range of 0.75 - 0.63, the adjustment of the learning rate is more refined, slowly dropping from 0.0035 to 0.002. In contrast, the AdaBelief algorithm (square markers) fixes the learning rate at 0.0048 after 400 steps and loses its adaptive ability; the RAdam algorithm (upper triangular markers) shows a simple linear decay characteristic, dropping from 0.01 to 0.005, and has a weak correlation with the improvement of system performance. By closely combining the learning rate adjustment with the change of the reward gain, this technical solution achieves a better adaptive optimization process. Especially in the interval with a drastic change in the reward gain (200 - 400 steps), the adjustment of the learning rate shows better followability and accuracy, fully reflecting the adaptive performance advantage of the algorithm.

[0183] In this embodiment, by introducing an adaptive step size and a warm-up mechanism, it is avoided that the parameter update is too large in the initial stage of training, resulting in instability, and the stability and convergence of model training are improved. By adopting a momentum estimation and bias correction mechanism, the problems of gradient sparsity and noise are overcome, making the parameter update smoother and accelerating the model convergence speed. By combining multi-scale parameter optimization and gradient correction, hierarchical dynamic update of parameters is realized, improving the model's learning ability and generalization performance for different scale features;

[0184] The control strategies of existing air water production systems usually adopt a fixed learning rate and a unified parameter update frequency, using the same update strategy for all control parameters, and being unable to make differential adjustments according to the response characteristics of different parameters, resulting in slow system response or overregulation. The fixed learning rate setting makes it difficult for the system to adaptively adjust the optimization step size during the training process, easily falling into local optimal solutions or having problems with unstable training. In this embodiment, by constructing an adaptive step size adjustment function and a periodic learning rate scheduling function, the dynamic adjustment of the learning rate is achieved, enabling the system to 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 initial 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, ensuring both the timely adjustment of fast parameters and the stability of slow parameters. In summary, this embodiment realizes the unity of short-term control accuracy and long-term operation stability, greatly improving the air water production efficiency, reducing energy consumption, and being significantly superior to traditional single update strategies.

[0185] An intelligent prediction and regulation system for air water production based on meteorological big data, comprising:

[0186] A first unit, configured to collect meteorological monitoring data through meteorological sensors, perform standardized 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] A second unit, configured to construct a transfer cross-domain learning network, add the pre-trained model knowledge in the refrigeration and air-conditioning field and the meteorological forecasting field to the air water production prediction task through a knowledge transfer algorithm, and fuse the expert knowledge of multiple source domains into a deep neural network by combining knowledge distillation;

[0188] A third unit, configured to train a deep neural network based on the enhanced feature space in combination with the backpropagation algorithm, output the predicted water production per hour within the next day, use the predicted water production as the input of the state space and configure the action space, construct a reward function according to the water production efficiency and the energy consumption ratio, and construct a deep reinforcement learning model based on the state space, the action space, and the reward function;

[0189] A fourth unit, configured to dynamically adjust the network structure of the deep reinforcement learning model through a meta-learning adaptive framework, 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 an optimization target;

[0190] The fifth unit is used for joint optimization through a multi-objective Bayesian optimizer, generating a Pareto optimal solution set and selecting optimal control parameters through a confidence-boundary-based dynamic decision-making algorithm, and dynamically adjusting the update step size of the deep reinforcement learning model based on the optimal control parameters to obtain an 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 having thereon computer-readable program instructions for performing 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 and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and 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