An artificial intelligence-based refrigeration station energy-saving control method and system

Abstract

The application discloses a kind of energy-saving control method and system of refrigeration station based on artificial intelligence, it is related to energy-saving control technical field, including collection multimodal data, optimization is carried out using PINN neural network, global thermal field map and airflow field are generated by QVC circuit and quantum variation optimization algorithm, global thermal field map is mapped to quantum hilbert space, cold load demand is predicted by NQ-GRU, and chaotic sequence is generated using Sine chaos mapping, and local thermal field feature is extracted, the action of equipment parameter is generated by DQN, the action of optimum equipment parameter is disturbed using chaotic sequence, and is converted into PLC instruction, and equipment operation instruction set is generated.The application is optimized by OFT circuit in combination with PINN neural network, and introduces quantum variation optimization algorithm, improves the precision and efficiency of refrigeration station energy-saving control, the output of model is disturbed by chaos mapping technology, and the flexibility of equipment operation and system adaptive capacity are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy-saving control, in particular to a refrigeration station energy-saving control method and system based on artificial intelligence. BACKGROUND

[0002] With the acceleration of industrialization and global warming, the energy consumption of the refrigeration system has become an important factor restricting the sustainable development of the industry. The traditional energy-saving control method of the refrigeration system usually relies on thermodynamic modeling, rule-based control strategy or experience formula driven optimization means. Artificial intelligence, especially deep learning and reinforcement learning, shows great potential in complex system modeling, prediction and control. PINN neural network, GRU and other methods are gradually applied to thermal system modeling and control, providing higher prediction accuracy and better control strategy for the refrigeration system. At the same time, quantum computing and chaotic optimization technology are gradually introduced to solve the bottleneck problems of high-dimensional multi-variable coupling modeling difficulty and local optimal trap.

[0003] The existing water-cooled air conditioner energy-saving optimization method still has defects. The existing PINN method is mostly based on pure data driving, lacks physical law embedding, and lacks collaborative application of quantum variational optimization and NQ-GRU, which is difficult to fully capture local thermal field characteristics and nonlinear dynamics, resulting in limited cold load prediction accuracy. The device parameter optimization mostly uses static action space, lacks deep integration with chaotic disturbance mechanism, is easy to fall into local optimum, and is difficult to respond to dynamic environmental changes in real time. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a refrigeration station energy-saving control method and system based on artificial intelligence, which solves the problem that the existing PINN method is mostly based on pure data driving, lacks physical law embedding, and lacks collaborative application of quantum variational optimization and NQ-GRU, which is difficult to fully capture local thermal field characteristics and nonlinear dynamics, resulting in limited cold load prediction accuracy. The device parameter optimization mostly uses static action space, lacks deep integration with chaotic disturbance mechanism, is easy to fall into local optimum, and is difficult to respond to dynamic environmental changes in real time.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present invention provides an energy-saving control method for a refrigeration station based on artificial intelligence, comprising: collecting multimodal data; generating an initial pressure field and temperature field through OFT circuits and linear mapping; optimizing the field using a PINN neural network; generating a global thermal field map and airflow field through QVC circuits and quantum variational optimization algorithms; mapping the global thermal field map to quantum Hilbert space; predicting cooling load demand through NQ-GRU; generating a chaotic sequence using Sine chaotic mapping and extracting local thermal field features; generating the action of equipment parameters through DQN; perturbing the action of the optimal equipment parameters using the chaotic sequence and converting it into PLC instructions; generating an equipment operation instruction set; allocating energy storage tasks through VMD decomposition method; generating an energy storage allocation instruction set; executing the instruction set and setting an alarm mechanism; and storing the collected and analyzed multimodal data.

[0008] As a preferred embodiment of the AI-based energy-saving control method for refrigeration stations described in this invention, the following steps are included: collecting multimodal data, generating initial pressure and temperature fields through OFT circuits and linear mapping, optimizing using a PINN neural network, generating a global thermal field map and airflow field through QVC circuits and quantum variational optimization algorithms, mapping the global thermal field map to quantum Hilbert space, and predicting cooling load demand using NQ-GRU.

[0009] Collect multimodal data, including temperature and humidity, wind speed, cooling load, chilled water flow rate, supply and return water temperature difference, chiller unit power, cooling tower fan speed, water pump frequency, valve opening degree, and equipment power consumption data.

[0010] Initialize the virtual CFD computing domain, extract temperature, humidity, wind speed and cooling load data from the multimodal data, define the virtual grid space, and calculate the number of grid points;

[0011] Temperature, humidity, and wind speed data are extracted from multimodal data, and grid point coordinates are generated using K-medoids.

[0012] Based on grid point coordinates and multimodal data, the initial velocity field of each grid point is generated by RBF interpolation, and the normalized velocity field is generated by mean normalization.

[0013] Using IBM Qiskit, the normalized velocity field is mapped to the quantum state through amplitude encoding. Using QFT circuits, the frequency domain velocity components are measured. The initial pressure field and initial temperature field are generated through linear mapping and normalized using Min-Max normalization.

[0014] Construct a PINN neural network to calculate the mass conservation loss, momentum conservation loss, and energy conservation loss;

[0015] Extract cooling load and wind speed data from multimodal data, and calculate cooling load fluctuation and wind speed fluctuation respectively using wavelet transform;

[0016] The total PINN loss is generated by weighted summation of the mass conservation loss, momentum conservation loss, and energy conservation loss.

[0017] Initialize the Adam optimizer, iteratively calculate the total loss of PINN through forward propagation, calculate the gradient of the total loss of PINN with respect to the PINN parameters using the automatic differentiation method, update the PINN weights and biases using the Adam optimizer based on the gradient, monitor the relative change of the gradient norm using the relative change method, and stop optimization when the relative change no longer decreases.

[0018] The normalized velocity field, normalized initial pressure field, normalized initial temperature field, multimodal data, and grid point coordinates are input into the optimized PINN neural network, which outputs the optimized velocity field, pressure field, and temperature field.

[0019] Based on multimodal data, turbulent viscosity parameters and thermal conductivity coefficients were calculated using the RANS method.

[0020] Based on the thermal conductivity coefficient and turbulent viscosity coefficient, a QVC circuit is constructed to encode the optimized velocity field and temperature field into the initial quantum state;

[0021] By using a quantum variational optimization algorithm and embedding CFD conservation loss, the QVC parameters are updated. The update stops when the maximum number of iterations is reached, and the local thermal field and airflow field are output.

[0022] The local thermal field and airflow field are stitched together using the kd-tree algorithm, and then smoothed by Gaussian filtering to generate a global thermal field map and airflow field.

[0023] A time series is constructed from the global thermal field map using a sliding window, converted into quantum states through angle encoding, mapped to quantum Hilbert space, entangled state features are generated through QVFE, and dimensionality is reduced through quantum principal component analysis to generate an enhanced feature subset.

[0024] An NQ-GRU model is constructed, which optimizes the update gate and reset gate through a quantum gating mechanism based on an enhanced feature subset. The hidden state at the current time is calculated and mapped through a fully connected layer to output the predicted cold load demand at future time.

[0025] As a preferred embodiment of the AI-based energy-saving control method for refrigeration stations described in this invention, the method involves: generating a chaotic sequence using Sine chaotic mapping, extracting local thermal field features, generating equipment parameter actions through DQN, perturbing the optimal equipment parameter actions using the chaotic sequence, and converting these perturbations into PLC instructions to generate an equipment operation instruction set, including:

[0026] Based on the analysis of the characteristics of chaotic sequences, an initial value was set, and the Sine chaotic mapping formula was used for iterative calculation until the maximum number of calculations was reached. All the calculation results were then concatenated into a chaotic sequence.

[0027] Based on the calculation results of the variance of the global thermal field map and the fluctuation range of cooling load demand, the number of sub-regions N is determined.

[0028] Based on chaotic sequences, the center point coordinates of N sub-regions are generated through polynomial interpolation;

[0029] Local thermal field features of N sub-regions are extracted, perturbed using Sine chaotic mapping, and the perturbed local thermal field features are concatenated into state vectors. The entropy value of each state vector is calculated using the entropy calculation method.

[0030] The local variance of each state vector is calculated using the biased sample variance method, and the dynamic exploration boundary is generated and concatenated into a dynamic exploration boundary set.

[0031] The exploration boundary set is grouped using k-means clustering, and an initial priority exploration region is set;

[0032] Calculate the coverage of the initial priority exploration area, set a coverage threshold based on historical operating data, filter areas with coverage greater than the coverage threshold, and output the priority exploration area and state vector;

[0033] Construct the MPC model and define the optimization objective function;

[0034] Based on the priority exploration region and state vector, the device parameters for predicting future time are output, and optimization is performed by minimizing the optimization objective function. Perturbation is applied through Sine chaotic mapping, and the perturbation results are integrated into a short-term control sequence.

[0035] Construct a DQN model and define the reward function;

[0036] The Q-value is maximized by maximizing the reward function. The global heat map and state vector are input into the DQN model, and the action with the optimal device parameters is output. The action with the optimal device parameters is perturbed by a chaotic sequence.

[0037] The disturbed actions and short-term control sequences are converted into PLC instructions to generate a set of equipment operation instructions.

[0038] As a preferred embodiment of the energy-saving control method for refrigeration stations based on artificial intelligence described in this invention, the step of allocating energy storage tasks through VMD decomposition to generate an energy storage allocation instruction set includes:

[0039] Based on the difference between the target power and the cooling load demand, U modal IMF components are output using the VMD decomposition method;

[0040] The center frequency of each IMF component is extracted by FFT. Based on the time scale mapping method, low frequency threshold and high frequency threshold are set respectively. The task group with center frequency greater than high frequency threshold is selected to generate high frequency mode, and the task group with center frequency less than low frequency threshold is selected to generate low frequency mode.

[0041] High-frequency modal tasks are assigned to supercapacitors, and low-frequency modal tasks are assigned to batteries to generate an energy storage task allocation scheme. This scheme is then converted into PLC instructions to generate an energy storage allocation instruction set.

[0042] As a preferred embodiment of the energy-saving control method for refrigeration stations based on artificial intelligence described in this invention, the step of executing the instruction set and setting the alarm mechanism includes:

[0043] The PLC controller executes the equipment operation instruction set and energy storage distribution instruction set, collects multimodal data after execution, and sets up an alarm mechanism based on the multimodal data after execution.

[0044] As a preferred embodiment of the energy-saving control method for refrigeration stations based on artificial intelligence described in this invention, the step of preprocessing the multimodal data includes:

[0045] Multimodal data is decomposed using EMD decomposition, and the decomposition results are smoothed and denoised using Kalman filtering. The denoised multimodal data is then normalized using Z-score and timestamps are added.

[0046] As a preferred embodiment of the AI-based energy-saving control method for refrigeration stations described in this invention, the storage, collection, and analysis of the generated multimodal data includes:

[0047] The collected multimodal data, along with the analysis-generated predictions, equipment operation instruction sets, and energy storage allocation instruction sets, are used to generate a complete assessment report through data visualization, which is then stored in a central database.

[0048] Secondly, the present invention provides an energy-saving control system for a refrigeration station based on artificial intelligence, comprising,

[0049] The collection and optimization module is used to collect multimodal data, generate initial pressure and temperature fields through OFT circuits and linear mapping, optimize them using PINN neural networks, generate global thermal field maps and airflow fields through QVC circuits and quantum variational optimization algorithms, map the global thermal field maps to quantum Hilbert space, and predict cooling load demand through NQ-GRU.

[0050] The equipment prediction module is used to generate chaotic sequences using Sine chaotic mapping, extract local thermal field features, generate actions for equipment parameters through DQN, perturb the actions of the optimal equipment parameters using chaotic sequences, and convert them into PLC instructions to generate a set of equipment operation instructions.

[0051] The energy storage allocation module is used to allocate energy storage tasks using the VMD decomposition method and generate an energy storage allocation instruction set.

[0052] The execution storage module is used to execute instruction sets, set alarm mechanisms, and store the multimodal data generated during collection and analysis.

[0053] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the energy-saving control method for refrigeration stations based on artificial intelligence as described in the first aspect of the present invention.

[0054] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the energy-saving control method for refrigeration stations based on artificial intelligence as described in the first aspect of the present invention.

[0055] The beneficial effects of this invention are as follows: By collecting multimodal data and optimizing it using a PINN neural network, this invention generates a global thermal field map and airflow field using a QVC circuit and a quantum variational optimization algorithm. The global thermal field map is then mapped to a quantum Hilbert space. Cooling load demand is predicted using an NQ-GRU, a chaotic sequence is generated using a Sine chaotic mapping, and local thermal field features are extracted. The actions of equipment parameters are generated using a DQN, and the chaotic sequence is used to perturb the actions of the optimal equipment parameters, which are then converted into PLC instructions to generate a set of equipment operation instructions. This improves the accuracy and efficiency of energy-saving control in refrigeration stations and enhances the flexibility of equipment operation and the system's adaptive capabilities. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Fig. 1 This is a flowchart of the energy-saving control method for a refrigeration station based on artificial intelligence in Example 1.

[0058] Fig. 2 This is a schematic diagram of the AI-based energy-saving control system for a refrigeration station in Example 1.

[0059] Fig. 3 This is a schematic diagram of the parameters and operation of the chaotic disturbance device in Example 1.

[0060] Fig. 4 This is a schematic diagram of the quantum optimization module in Example 1. Detailed Implementation

[0061] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0062] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0063] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0064] Example 1, referring to Figs. 1 to 4 This is the first embodiment of the present invention, which provides an energy-saving control method for a refrigeration station based on artificial intelligence, including the following steps:

[0065] S1. Collect multimodal data, generate initial pressure and temperature fields through OFT circuits and linear mapping, optimize using PINN neural network, generate global thermal field map and airflow field through QVC circuits and quantum variational optimization algorithm, map global thermal field map to quantum Hilbert space, and predict cooling load demand through NQ-GRU.

[0066] Specifically, collecting multimodal data includes:

[0067] Multimodal data is collected and preprocessed using smart sensors;

[0068] The intelligent sensors include temperature and humidity sensors, wind speed sensors, cooling load sensors, flow rate sensors, temperature difference sensors, power sensors, rotation speed sensors, frequency sensors, rotary position sensors, and power consumption sensors.

[0069] The multimodal data includes temperature and humidity, wind speed, cooling load, chilled water flow rate, supply and return water temperature difference, chiller unit power, cooling tower fan speed, water pump frequency, valve opening degree, and equipment power consumption data.

[0070] The preprocessing includes decomposing the multimodal data using EMD decomposition, smoothing and denoising the decomposition results using Kalman filtering, normalizing the denoised multimodal data using Z-score, and adding timestamps.

[0071] By using multiple types of sensors, such as temperature, humidity, wind speed, flow rate, and power, and combining EMD decomposition, Kalman filtering, and Z-score normalization, we ensure the temporal consistency, stability, and universal scale representation of the data, providing a clean and usable input source for subsequent models and improving computational stability.

[0072] Furthermore, initial pressure and temperature fields are generated using OFT circuits and linear mapping, optimized using a PINN neural network, and a global thermal field map and airflow field are generated using QVC circuits and quantum variational optimization algorithms. The global thermal field map is mapped to quantum Hilbert space, and cooling load demand is predicted using NQ-GRU, including:

[0073] Initialize the virtual CFD computation domain, extract temperature, humidity, wind speed, and cooling load data from the multimodal data, define the virtual grid space, and calculate the number of grid points using the following formula:

[0074] V = α·Q + β·v,

[0075]

[0076] Where, N grid Where is the number of grid points, Q is the cooling load, v is the wind speed, α and β are coefficients for cooling load and wind speed respectively (set based on empirical analysis), V is the effective computational volume, and R is the effective computational volume. 3 The grid resolution (obtained experimentally);

[0077] Temperature, humidity, and wind speed data are extracted from multimodal data, and grid point coordinates are generated using K-medoids.

[0078] Based on grid point coordinates and multimodal data, the initial velocity field of each grid point is generated by RBF interpolation, and the normalized velocity field is generated by mean normalization.

[0079] Using IBM Qiskit, the normalized velocity field is mapped to the quantum state through amplitude encoding, and the frequency domain velocity component is measured using QFT circuitry.

[0080] Based on the frequency domain velocity component, the initial pressure field and initial temperature field are generated by linear mapping and normalized using Min-Max normalization.

[0081] Construct a PINN neural network, including an input layer, hidden layers, and an output layer;

[0082] Based on the normalized velocity field, the mass conservation loss is calculated by the finite difference method; based on the chiller power and cooling tower fan speed data in the multimodal data, the momentum conservation loss is calculated by the Navier-Stokes equations; and based on the normalized initial temperature field and the cooling load data of the multimodal data, the energy conservation loss is calculated by the heat conduction equations.

[0083] Extract cooling load and wind speed data from multimodal data, and calculate cooling load fluctuation and wind speed fluctuation respectively using wavelet transform;

[0084] The mass conservation loss, momentum conservation loss, and energy conservation loss are weighted and summed (the weights of mass conservation loss, momentum conservation loss, and energy conservation loss are initialized using a uniform initialization method; thresholds for cooling load fluctuation and wind speed fluctuation are set according to empirical rules; cooling load fluctuation is selected if it exceeds the cooling load fluctuation threshold or wind speed fluctuation is exceeded; the weights of energy conservation and momentum conservation are adjusted; the total weight of the three is always 1) to generate the PINN total loss.

[0085] Initialize the Adam optimizer, iteratively calculate the total loss of PINN through forward propagation, calculate the gradient of the total loss of PINN with respect to the PINN parameters using the automatic differentiation method, update the PINN weights and biases using the Adam optimizer based on the gradient, monitor the relative change of the gradient norm using the relative change method, and stop optimization when the relative change no longer decreases.

[0086] The normalized velocity field, normalized initial pressure field, normalized initial temperature field, multimodal data, and grid point coordinates are input into the optimized PINN neural network, which outputs the optimized velocity field, pressure field, and temperature field.

[0087] Based on multimodal data, turbulent viscosity parameters and thermal conductivity coefficients were calculated using the RANS method.

[0088] Based on the thermal conductivity and turbulent viscosity, a QVC circuit is constructed, including qubits, RY, RZ, and CNOT gates, to encode the optimized velocity and temperature fields into initial quantum states, as shown in the following formula:

[0089]

[0090] Where Z is the normalization factor, v ' and k are the thermal conductivity coefficient and turbulent viscosity coefficient, respectively, |ψ> is the initial quantum state, M is the number of grid points, and a i Let x be the quantum state amplitude at the i-th grid point. i and u i Let T be the coordinates of the i-th grid point and the magnitude of the optimized velocity field (calculated based on the L2 norm), respectively. i The optimized temperature field for the i-th grid point;

[0091] By using a quantum variational optimization algorithm and embedding CFD conservation loss, the QVC parameters are updated. The update stops when the maximum number of iterations is reached (based on cross-validation settings), and the local thermal field and airflow field are output.

[0092] The local thermal field and airflow field are stitched together using the kd-tree algorithm, and then smoothed by Gaussian filtering to generate a global thermal field map and airflow field.

[0093] A time series is constructed from the global thermal field map using a sliding window, converted into quantum states through angle encoding, mapped to quantum Hilbert space, entangled state features are generated through QVFE, and dimensionality is reduced through quantum principal component analysis to generate an enhanced feature subset.

[0094] Construct an NQ-GRU model, including an input layer, a GRU layer, a fully connected layer, and an output layer;

[0095] Physical simulation data is collected through a digital twin platform, mapped to quantum Hilbert space, and a subset of simulation-enhanced features is generated. This subset is then input into the NQ-GRU model for training and optimized using the MSE loss function.

[0096] Based on the enhanced feature subset, the update gate and reset gate are optimized through a quantum gating mechanism to calculate the hidden state at the current time.

[0097] By mapping the current hidden state to the prediction target through a fully connected layer, the cold load demand for the predicted future time is output.

[0098] By mapping the velocity field to the quantum frequency domain using amplitude encoding and QFT circuits, and then deriving the pressure and temperature fields through linear mapping, the expressive power of frequency domain features in physical modeling is greatly enhanced. This helps to reveal the microscopic mechanisms of high-frequency disturbances and energy consumption changes in the flow field. By introducing three types of physical loss functions—mass conservation, momentum conservation, and energy conservation—and adaptively adjusting the loss weights based on cooling load and wind speed fluctuations, the physical consistency and robustness of the model under dynamic boundary conditions are effectively enhanced. Using QVC circuits and variational optimization algorithms, the thermal conductivity coefficient and turbulent viscosity information are encoded through adjustable electrical... The airflow field is generated by gate optimization, achieving scene reconstruction that closely conforms to physical laws and providing high-precision input data for cold load prediction. After the global thermal field map is converted into a time series, it is mapped to quantum states through angle encoding. Then, QVFE is used to extract entanglement features, and quantum PCA is combined for dimensionality reduction, which effectively compresses the feature dimension while retaining key information, improving learning efficiency and generalization ability. A quantum gating mechanism is introduced to dynamically adjust the weights of the update gate and the reset gate, enabling the model to have better memory stability and nonlinear modeling ability when learning complex time series patterns, further improving prediction accuracy and engineering usability.

[0099] S2. Use Sine chaotic mapping to generate chaotic sequences and extract local thermal field features. Generate device parameter actions using DQN, perturb the optimal device parameter actions using the chaotic sequences, and convert them into PLC instructions to generate a device operation instruction set, including:

[0100] Based on the analysis of chaotic sequence characteristics, an initial value was set, and iterative calculations were performed using the Sine chaotic mapping formula until the maximum number of calculations was reached (based on the actual application scenario). All calculation results were then concatenated into a chaotic sequence, with the formula as follows:

[0101]

[0102] Where, q c is the initial value for the iteration of the chaotic mapping formula, c is the iteration index, b is the control parameter (set according to the maximum chaos requirement), and π is the value of pi.

[0103] Based on the variance of the global thermal field map (calculated using the unbiased sample variance formula) and the fluctuation range of cooling load demand (the difference between the maximum and minimum values), the number of sub-regions is determined using the following formula:

[0104] N = τ·e + γ·ΔL,

[0105] Where N is the number of sub-regions, e is the variance of the global thermal field map, ΔL is the fluctuation range of cooling load demand, and τ and γ are the weighting coefficients of the variance of the global thermal field map and the fluctuation range of cooling load demand, respectively (based on the initial setting of uniform distribution).

[0106] Based on chaotic sequences, the center point coordinates of N sub-regions are generated through polynomial interpolation;

[0107] The formula for extracting the local thermal field features of N sub-regions is as follows:

[0108]

[0109] Where F represents the local thermal field characteristics, This represents the average temperature of the sub-region (obtained by summing and averaging the temperatures of the sub-regions). The temperature gradient is calculated by convolving the temperature field in the global thermal field map of the sub-region using the Sobel operator, and A is the cooling load demand.

[0110] The local thermal field features are perturbed using Sine chaotic mapping, and the perturbed local thermal field features are concatenated into a state vector. The entropy value of each state vector is calculated using the entropy calculation method.

[0111] The local variance of each state vector is calculated using the biased sample variance method to generate a dynamic exploration boundary. These boundaries are then concatenated to form a dynamic exploration boundary set, as shown in the formula:

[0112]

[0113] Where B represents the dynamic exploration boundary. Let t be the mean of the state vector. y,r y is the critical value of the distribution (obtained based on the t-distribution table), y is the degree of freedom (based on the t-distribution variation strategy, combined with the entropy value of each state vector, and set by the Sigmoid function), r is the confidence level (obtained based on historical operating condition analysis), and θ is the standard deviation of the state vector (calculated based on the square root method).

[0114] The exploration boundary set is grouped using k-means clustering. The entropy values ​​of each state vector after grouping are sorted in descending order. The j regions with the highest entropy values ​​in the grouping results (based on the cumulative contribution rate of entropy values) are selected and set as the initial priority exploration regions.

[0115] Calculate the coverage of the initial priority exploration area (the ratio of the exploration boundary set to the initial priority exploration area), set a coverage threshold based on historical working data, filter areas with coverage greater than the coverage threshold, and output the priority exploration area and state vector.

[0116] Construct the MPC model, including the input layer, design layer, and output layer;

[0117] The MPC model was trained using a simulation dataset obtained through the COMSOL simulation platform.

[0118] Based on the priority exploration of the region and state vector, the MPC model is used to output the equipment parameters for predicting future times, including chiller load, cooling tower fan speed, water pump frequency and valve opening.

[0119] Based on the average entropy value of the priority exploration region, the optimization objective function is defined as follows:

[0120]

[0121] Where J is the value of the objective function. Here, E is the sum of the power consumption of all devices, where E is the weighting factor. max The maximum sum of power consumption of all devices is given by PMV, where ΔT” and ΔP are the temperature deviation and pressure deviation (the difference between predicted and actual temperatures and the difference between optimized and actual pressure fields, respectively). sc To predict PMV values ​​(based on PMV model predictions), H avg The average entropy value of the priority region;

[0122] The prediction results of the MPC model are optimized by minimizing the objective function, and the perturbation is applied by the Sine chaotic mapping to integrate the perturbation results into a short-term control sequence.

[0123] Construct a DQN model, including an input layer, hidden layers, and an output layer;

[0124] Simulation data was obtained using Gazebo to train the DQN model;

[0125] Define the reward function, with the following formula:

[0126]

[0127] Where R is the function reward value, and o is the coverage of the priority exploration area;

[0128] By maximizing the reward function to maximize the Q value, the global thermal map and state vector are input into the DQN model, and the action with the best device parameters is output.

[0129] Use chaotic sequences to perturb the actions of the optimal device parameters;

[0130] The disturbed actions and short-term control sequences are converted into PLC instructions to generate a set of equipment operation instructions.

[0131] By perturbing the device parameters and thermal field characteristics using Sine chaotic mapping, the local optimum trap in reinforcement learning is effectively avoided, helping the system escape overfitting regions and improving exploration efficiency and control accuracy. Local thermal field features extracted at the sub-region scale, combined with temperature gradients and load changes, make system regulation more targeted, responding to local abnormal states and improving the system's adaptability to environmental dynamics. Entropy combined with local variance is used to construct dynamic exploration boundaries, considering not only state complexity but also quantifying the range of uncertainty changes, enabling dynamic adjustment of the control strategy's tolerance to environmental disturbances and improving the system's robustness. Through entropy-based clustering and coverage screening strategies, resource allocation and device adjustment are performed not only in high-value areas, avoiding redundant control across the entire space, but the optimized objective function integrates device power consumption, temperature and pressure deviations, PMV human comfort index, and information entropy indicators, reflecting the joint optimization idea of ​​multiple objectives and constraints. This not only improves energy utilization efficiency but also takes into account operational stability and terminal comfort.

[0132] S3. Energy storage task allocation is performed using the VMD decomposition method, generating an energy storage allocation instruction set, including:

[0133] Based on the difference between the target power (obtained by weighted summation of mechanical and thermal power) and the cooling load demand, U modal IMF components (set based on actual demand) are output using the VMD decomposition method, as shown in the formula:

[0134]

[0135] p 2 =m·c p ·ΔT 1 ,

[0136] Where, p 1 and p 2 These are mechanical power and thermal power, ΔP * h Q is the pressure difference (obtained based on the pressure field). ' h η is the cold water flow rate. h The efficiency is denoted as ΔT (based on actual equipment testing), m is the mass flow rate (based on the product of cold water flow rate and water density), and ΔT is the mass flow rate. 1 For the supply and return water temperature difference, c p Where is the specific heat capacity of water, and h is the index of the equipment;

[0137] The center frequency of each IMF component is extracted by FFT. Based on the time scale mapping method, low frequency threshold and high frequency threshold are set respectively. The task group with center frequency greater than high frequency threshold is selected to generate high frequency mode, and the task group with center frequency less than low frequency threshold is selected to generate low frequency mode.

[0138] High-frequency modal tasks are assigned to supercapacitors, and low-frequency modal tasks are assigned to batteries to generate an energy storage task allocation scheme. This scheme is then converted into PLC instructions to generate an energy storage allocation instruction set.

[0139] By using the difference between the target power and the cooling load as a non-stationary signal input to the VMD framework, multi-scale fluctuation information of the system load can be effectively extracted. The center frequency of each mode is obtained through FFT, and the frequency threshold is set using a time-scale mapping method. This enables the system to distinguish between high-frequency and low-frequency tasks based on the speed of mode change, realizing frequency domain feature-driven task allocation. The frequency-based task allocation strategy not only improves the utilization efficiency of the energy storage unit, but also extends the service life of the equipment and optimizes the energy flow path.

[0140] S4. Execute the instruction set and set the alarm mechanism, and store the collected and analyzed multimodal data;

[0141] Specifically, execute the instruction set and set up an alarm mechanism, including:

[0142] The PLC controller executes equipment operation and energy storage distribution instruction sets, collects multimodal data after execution, and sets alarm thresholds (based on historical equipment operating status). If the executed multimodal data deviates from the alarm threshold... When Y is the fluctuation range, set based on statistical analysis, a low-level alarm is triggered, notifying staff to conduct an inspection;

[0143] If the multimodal data after execution deviates from the alarm threshold Y, a medium-level alarm will be triggered, notifying staff to make adjustments and perform maintenance.

[0144] If the multimodal data after execution deviates from the alarm threshold of 2Y, an advanced alarm will be triggered, and the system will be shut down immediately.

[0145] By executing operation and energy storage distribution instructions through PLC, not only is standardized control of processes such as refrigeration and energy supply achieved, but data and behavior linkage between different modules is also effectively integrated. This supports rapid response to environmental disturbances and predictive maintenance. Statistical analysis methods are used to set floating thresholds, significantly enhancing the system's adaptability and early warning capabilities. Alarm sensitivity is dynamically adjusted to reduce false alarms and missed alarms, ensuring the accuracy and timeliness of alarms. By classifying alarms into three levels, the system's fault tolerance is improved, and the allocation of operation and maintenance resources is optimized.

[0146] Furthermore, the storage, collection, and analysis of the resulting multimodal data includes:

[0147] The collected multimodal data, along with the analysis-generated predictions, equipment operation instruction sets, and energy storage allocation instruction sets, are used to generate a complete assessment report through data visualization, which is then stored in a central database.

[0148] By generating data visualization and evaluation reports, multi-dimensional information is presented to users, helping them quickly understand the system's operating status and optimization results. This provides data-driven decision support for management, helping to improve the scientific nature and accuracy of management. By storing data in a central database, it provides data support for global optimization and improves the efficiency of collaborative work.

[0149] This embodiment also provides an energy-saving control system for refrigeration stations based on artificial intelligence, including:

[0150] The collection and optimization module is used to collect multimodal data, generate initial pressure and temperature fields through OFT circuits and linear mapping, optimize them using PINN neural networks, generate global thermal field maps and airflow fields through QVC circuits and quantum variational optimization algorithms, map the global thermal field maps to quantum Hilbert space, and predict cooling load demand through NQ-GRU.

[0151] The equipment prediction module is used to generate chaotic sequences using Sine chaotic mapping, extract local thermal field features, generate actions for equipment parameters through DQN, perturb the actions of the optimal equipment parameters using chaotic sequences, and convert them into PLC instructions to generate a set of equipment operation instructions.

[0152] The energy storage allocation module is used to allocate energy storage tasks using the VMD decomposition method and generate an energy storage allocation instruction set.

[0153] The execution storage module is used to execute instruction sets, set alarm mechanisms, and store the multimodal data generated during collection and analysis.

[0154] This embodiment also provides a computer device applicable to the energy-saving control method for refrigeration stations based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the energy-saving control method for refrigeration stations based on artificial intelligence as proposed in the above embodiment.

[0155] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0156] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the energy-saving control method for a cooling station based on artificial intelligence as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0157] In summary, this invention collects multimodal data, optimizes it using a PINN neural network, generates a global thermal field map and airflow field using a QVC circuit and a quantum variational optimization algorithm, maps the global thermal field map to a quantum Hilbert space, predicts cooling load demand using an NQ-GRU, generates a chaotic sequence using a Sine chaotic mapping and extracts local thermal field features, generates equipment parameter actions using a DQN, perturbs the optimal equipment parameter actions using the chaotic sequence, and converts them into PLC instructions to generate an equipment operation instruction set; thus improving the accuracy and efficiency of energy-saving control in refrigeration stations, enhancing the flexibility of equipment operation and the system's adaptive capabilities.

[0158] 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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An energy-saving control method for refrigeration stations based on artificial intelligence, characterized in that: include, Collect multimodal data, including temperature and humidity, wind speed, cooling load, chilled water flow rate, supply and return water temperature difference, chiller unit power, cooling tower fan speed, water pump frequency, valve opening degree, and equipment power consumption data. Initialize the virtual CFD computing domain, extract temperature, humidity, wind speed and cooling load data from the multimodal data, define the virtual grid space, and calculate the number of grid points; Temperature, humidity, and wind speed data are extracted from multimodal data, and grid point coordinates are generated using K-medoids. Based on grid point coordinates and multimodal data, the initial velocity field of each grid point is generated by RBF interpolation, and the normalized velocity field is generated by mean normalization. Using IBM Qiskit, the normalized velocity field is mapped to the quantum state through amplitude encoding. Using QFT circuits, the frequency domain velocity components are measured. The initial pressure field and initial temperature field are generated through linear mapping and normalized using Min-Max normalization. Construct a PINN neural network to calculate the mass conservation loss, momentum conservation loss, and energy conservation loss; Extract cooling load and wind speed data from multimodal data, and calculate cooling load fluctuation and wind speed fluctuation respectively using wavelet transform; The total PINN loss is generated by weighted summation of the mass conservation loss, momentum conservation loss, and energy conservation loss. Initialize the Adam optimizer, iteratively calculate the total loss of PINN through forward propagation, calculate the gradient of the total loss of PINN with respect to the PINN parameters using the automatic differentiation method, update the PINN weights and biases using the Adam optimizer based on the gradient, monitor the relative change of the gradient norm using the relative change method, and stop optimization when the relative change no longer decreases. The normalized velocity field, normalized initial pressure field, normalized initial temperature field, multimodal data, and grid point coordinates are input into the optimized PINN neural network, which outputs the optimized velocity field, pressure field, and temperature field. Based on multimodal data, turbulent viscosity parameters and thermal conductivity coefficients were calculated using the RANS method. Based on the thermal conductivity coefficient and turbulent viscosity coefficient, a QVC circuit is constructed to encode the optimized velocity field and temperature field into the initial quantum state; By using a quantum variational optimization algorithm and embedding CFD conservation loss, the QVC parameters are updated. The update stops when the maximum number of iterations is reached, and the local thermal field and airflow field are output. The local thermal field and airflow field are stitched together using the kd-tree algorithm, and then smoothed by Gaussian filtering to generate a global thermal field map and airflow field. A time series is constructed from the global thermal field map using a sliding window, converted into quantum states through angle encoding, mapped to quantum Hilbert space, entangled state features are generated through QVFE, and dimensionality is reduced through quantum principal component analysis to generate an enhanced feature subset. An NQ-GRU model is constructed, which is based on an enhanced feature subset. The update gate and reset gate are optimized through a quantum gating mechanism to calculate the hidden state at the current time. The hidden state is then mapped through a fully connected layer to output the predicted cold load demand at future times. Based on the analysis of the characteristics of chaotic sequences, an initial value was set, and the Sine chaotic mapping formula was used for iterative calculation until the maximum number of calculations was reached. All the calculation results were then concatenated into a chaotic sequence. Based on the calculation results of the variance of the global thermal field map and the fluctuation range of cooling load demand, the number of sub-regions N is determined. Based on chaotic sequences, the center point coordinates of N sub-regions are generated through polynomial interpolation; Local thermal field features of N sub-regions are extracted, perturbed using Sine chaotic mapping, and the perturbed local thermal field features are concatenated into state vectors. The entropy value of each state vector is calculated using the entropy calculation method. The local variance of each state vector is calculated using the biased sample variance method, and the dynamic exploration boundary is generated and concatenated into a dynamic exploration boundary set. The exploration boundary set is grouped using k-means clustering, and an initial priority exploration region is set; Calculate the coverage of the initial priority exploration area, set a coverage threshold based on historical operating data, filter areas with coverage greater than the coverage threshold, and output the priority exploration area and state vector; Construct the MPC model and define the optimization objective function; Based on the priority exploration region and state vector, the device parameters for predicting future time are output, and optimization is performed by minimizing the optimization objective function. Perturbation is applied through Sine chaotic mapping, and the perturbation results are integrated into a short-term control sequence. Construct a DQN model and define the reward function; By maximizing the reward function to maximize the Q value, the global thermal field map and state vector are input into the DQN model, and the action of the optimal equipment parameters is output. The action of the optimal equipment parameters is perturbed by the chaotic sequence, and the perturbed action and short-term control sequence are converted into PLC instructions to generate the equipment operation instruction set. Based on the difference between the target power and the cooling load demand, U modal IMF components are output using the VMD decomposition method; The center frequency of each IMF component is extracted by FFT. Based on the time scale mapping method, low frequency threshold and high frequency threshold are set respectively. The task group with center frequency greater than high frequency threshold is selected to generate high frequency mode, and the task group with center frequency less than low frequency threshold is selected to generate low frequency mode. High-frequency modal tasks are assigned to supercapacitors, and low-frequency modal task groups are assigned to batteries to generate an energy storage task allocation scheme, which is then converted into PLC instructions to generate an energy storage allocation instruction set. Execute instruction sets and set alarm mechanisms, and store and analyze the generated multimodal data.

2. The energy-saving control method for refrigeration stations based on artificial intelligence as described in claim 1, characterized in that: The execution instruction set and alarm mechanism include: The PLC controller executes the equipment operation instruction set and energy storage distribution instruction set, collects multimodal data after execution, and sets up an alarm mechanism based on the multimodal data after execution.

3. The energy-saving control method for refrigeration stations based on artificial intelligence as described in claim 1, characterized in that: The multimodal data is first preprocessed, including: Multimodal data is decomposed using EMD decomposition, and the decomposition results are smoothed and denoised using Kalman filtering. The denoised multimodal data is then normalized using Z-score and timestamps are added.

4. The energy-saving control method for refrigeration stations based on artificial intelligence as described in claim 1, characterized in that: The multimodal data collected and analyzed includes: The collected multimodal data, along with the analysis-generated predictions, equipment operation instruction sets, and energy storage allocation instruction sets, are used to generate a complete assessment report through data visualization, which is then stored in a central database.

5. An energy-saving control system for a refrigeration station based on artificial intelligence, based on the energy-saving control method for a refrigeration station based on artificial intelligence as described in any one of claims 1 to 4, characterized in that: include, The collection and optimization module is used to collect multimodal data, generate initial pressure and temperature fields through OFT circuits and linear mapping, optimize them using PINN neural networks, generate global thermal field maps and airflow fields through QVC circuits and quantum variational optimization algorithms, map the global thermal field maps to quantum Hilbert space, and predict cooling load demand through NQ-GRU. The equipment prediction module is used to generate chaotic sequences using Sine chaotic mapping, extract local thermal field features, generate actions for equipment parameters through DQN, perturb the actions of the optimal equipment parameters using chaotic sequences, and convert them into PLC instructions to generate a set of equipment operation instructions. The energy storage allocation module is used to allocate energy storage tasks using the VMD decomposition method and generate an energy storage allocation instruction set. The execution storage module is used to execute instruction sets, set alarm mechanisms, and store the multimodal data generated during collection and analysis.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the energy-saving control method for refrigeration stations based on artificial intelligence as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the energy-saving control method for refrigeration stations based on artificial intelligence as described in any one of claims 1 to 4.

Images