Multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method

Abstract

The invention discloses a multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method. The method comprises the following steps: constructing a multi-source heterogeneous parameter acquisition system and preprocessing data; fusing the parameter spatio-temporal features through a multi-head spatio-temporal feature extraction module embedded with an attention mechanism; establishing a load prediction model integrating a long-short-term memory network and thermodynamic physical constraints; on the basis of a model predictive control framework, a future control instruction is optimized under the condition that comfort and equipment safety constraints are met; an online learning and updating mechanism based on prediction error triggering is introduced, and the model weight is dynamically and finely adjusted to adapt to system characteristic changes. According to the technical scheme, the load prediction precision and the control response speed under the complex working condition are remarkably improved, the real-time optimization of the system energy efficiency is achieved, and the long-term operation stability is guaranteed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of heat pump system control, and particularly relates to a multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method. BACKGROUND

[0002] With the deep integration of intelligent regulation and building energy-saving technology, central air conditioning systems are rapidly developing towards high energy efficiency and intelligence. As a new type of high-efficiency heat pump technology, air suspension heat pumps have broad application prospects in green buildings and low-carbon energy systems due to their advantages such as oil-free lubrication, low friction loss and wide operating condition adaptability. However, its efficient operation is highly dependent on accurate load prediction and dynamic energy-saving control strategies, which puts higher requirements on the fusion ability of multi-source operating parameters, the accuracy of nonlinear dynamic modeling and real-time optimization under physical constraints. Especially under the coupling action of complex weather conditions, building thermal inertia changes and user behavior disturbances, traditional control methods are difficult to accurately predict and energy-efficiently adjust the load demand of air suspension heat pump systems.

[0003] Among them, multi-parameter fusion AI load prediction is the key link to improve the energy-saving control performance of air suspension heat pumps. Although existing technologies attempt to introduce long short-term memory (LSTM), graph neural networks or Kalman filtering methods for load modeling, they generally lack consideration of the unique operating boundaries of air suspension heat pumps, such as compressor speed limits, refrigerant flow constraints, and thermodynamic cycle efficiency inflection points. At the same time, most models simply concatenate or linearly weight heterogeneous data such as environmental temperature and humidity, device status, and historical load, lacking deep mining of nonlinear coupling relationships between parameters and time-space cross-scale characteristics, resulting in significant prediction deviation under extreme operating conditions or rapid load change scenarios. In addition, existing solutions generally lack online correction mechanisms based on actual operation feedback, and cannot dynamically adjust model weights or control strategies based on measured energy consumption and prediction errors, resulting in control instructions lagging behind the actual system demand.

[0004] Therefore, there is an urgent need for a multi-parameter fusion AI load prediction energy-saving control method that takes into account the physical interpretability, nonlinear modeling capability and engineering realizability, which can deeply integrate the thermodynamic characteristics and operating constraints of air suspension heat pumps, build a lightweight and online learning capable prediction-control closed loop, thereby maximizing system energy efficiency while ensuring comfort, and solving the technical problems of inaccurate prediction, rigid regulation and insufficient release of energy-saving potential under complex operating conditions in existing technologies. SUMMARY

[0005] The present application aims to provide a multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method, which can effectively solve the technical problems of the existing load prediction method mentioned in the background art, such as insufficient consideration of the unique operating boundary of the air suspension heat pump, insufficient mining of the nonlinear coupling relationship and cross-scale characteristics between multi-source heterogeneous parameters, and lack of online correction mechanism based on actual operation feedback, resulting in large prediction deviation, control instruction lag, and insufficient release of system energy efficiency under complex working conditions.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present application is: A multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method, comprising the following steps: Step S1, multi-source heterogeneous data acquisition and preprocessing: Through the sensor network deployed in the air suspension heat pump system, real-time acquisition of environmental parameters, equipment operating parameters and historical load data is carried out, and the original data is preprocessed by filtering and outlier elimination; Step S2, spatio-temporal feature fusion and representation: The preprocessed multi-source parameters are input into a multi-head spatio-temporal feature extraction module embedded with an attention mechanism, which extracts the local features of the time series of each parameter through a time convolution network, and models the dynamic correlation between parameters through a graph attention network, finally generating a fused high-dimensional feature tensor; Step S3, load prediction with physical information constraint: The high-dimensional feature tensor is input into a load prediction model to obtain a load prediction value; the load prediction model integrates a data-driven long short-term memory network and a differentiable physical computation graph based on the first law of thermodynamics and system operation constraints, and the prediction value is the weighted sum of the preliminary prediction value of the long short-term memory network and the physical residual term; Step S4, real-time energy-saving optimization based on model predictive control: The load prediction sequence of multiple future periods is input into a rolling optimization model predictive controller, which takes the minimization of the total energy consumption of the system as the objective function, and is subject to hard constraints composed of the indoor temperature comfort interval, the compressor speed safety interval and the refrigerant pressure limit, to obtain the optimal control instruction sequence; Step S5, online learning and model adaptive update: After each control period ends, the error between the actual load and the predicted load is calculated; when the error exceeds the preset threshold for multiple consecutive periods, the online update mechanism is triggered, and the latest operating data is used to optimize and adjust part of the parameters of the load prediction model.

[0007] Preferably, the sensor network in step S1 includes high-precision temperature and humidity sensors, pressure sensors, and turbine flow meters, wherein the temperature measurement accuracy of the temperature and humidity sensors is ±0.2℃, the humidity measurement accuracy is ±2%RH, the pressure sensor has a range of 0-3MPa and an accuracy of ±0.5% full scale, and the turbine flow meter has an accuracy of ±1% reading. All sensor data is transmitted to the edge computing gateway through an industrial Ethernet network, and the transmission delay is less than 50 milliseconds.

[0008] Preferably, the multi-head spatio-temporal feature extraction module in step S2 includes 4 independent attention heads, each head has a hidden layer dimension of 64, the convolution kernel size of the temporal convolution network is 3, and the adjacency matrix of the graph attention network is dynamically constructed according to the Pearson correlation coefficient between parameters, and the correlation coefficient threshold is set to 0.7.

[0009] Preferably, the long short-term memory network in step S3 has 2 layers, each layer has 128 hidden units, and the loss rate is set to 0.1. The physical computational graph explicitly encodes the cubic relationship between compressor power and speed, and the product relationship between evaporator heat exchange capacity and refrigerant flow rate and temperature difference. The weight coefficient of the physical residual term is obtained through back propagation learning.

[0010] Preferably, the objective function of the model predictive controller in step S4 is expressed as the total energy consumption equal to the compressor power consumption plus the circulating water pump power consumption plus the fan power consumption, wherein the compressor power consumption is calculated through its power characteristic curve fitting formula, the solver uses the interior point method, the control period is 15 minutes, the prediction time domain is 8 cycles, and the control time domain is 4 cycles.

[0011] Preferably, the online updating mechanism in step S5 uses the small batch gradient descent algorithm, the batch size is fixed at 32, the learning rate is set to 0.001, the upper limit of the number of iterations for each update is 10 times, and the L2 norm of the model weight change is set to be less than 0.01 as an early stopping condition.

[0012] Preferably, the method further includes a running state evaluation and early warning mechanism based on real-time collected vibration data and current harmonic analysis, which uses a lightweight one-dimensional convolutional neural network to identify early abnormal states of aerostatic bearings. When the abnormal confidence exceeds 90%, an early warning is triggered, and the early warning information is input as a feature into the load prediction model to correct the prediction bias.

[0013] Preferably, the training data of the load prediction model covers at least two complete refrigeration and heating seasons, and the total amount of data is not less than 100,000. During training, the mean square error is used as the loss function, and the L2 regularization term is added with a regularization coefficient of 0.001. The Adam optimizer is used for training with an initial learning rate of 0.01.

[0014] Preferably, the method is deployed in a hybrid architecture composed of edge computing nodes and cloud data centers, the edge computing nodes are responsible for real-time data collection, feature fusion, prediction and control instruction generation of steps S1 to S4, and the calculation delay is less than 1 second; the cloud data center is responsible for model training and historical data management of step S5, and the edge and the cloud perform encrypted data synchronization through 4G or 5G network, and the daily data traffic is not more than 100MB.

[0015] Preferably, all the models involved in the method are deployed in the format of quantization to 16-bit floating point numbers.

[0016] Compared with the prior art, the present application has the following beneficial effects: 1. By constructing a deep fusion of multi-source heterogeneous parameter spatio-temporal features and embedding a differentiable physical constraint lightweight prediction model, the model expression ability is guaranteed while the deployability is considered, and the load prediction accuracy of the air suspension heat pump under complex working conditions is significantly improved; 2. By combining high-precision prediction with model predictive control with hard constraints, real-time closed-loop regulation and control with optimal energy efficiency are realized, and control lag is effectively avoided; 3. Through the lightweight online updating mechanism triggered based on the measured error, the model has the ability to continuously adapt to system performance drift; 4. The whole method adopts modular and low complexity design, and the core algorithm can be efficiently executed on resource-limited edge devices, which has good engineering realizability and popularization value, thereby fully releasing the energy saving potential of the air suspension heat pump system. BRIEF DESCRIPTION OF DRAWINGS

[0017] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the accompanying drawings: Figure 1 is the overall technical scheme architecture schematic diagram of the multi-parameter fusion air suspension heat pump AI load prediction energy saving control method proposed by the present application; Figure 2 is the flowchart of real-time energy saving optimization based on model predictive control in the method of the present application. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the present application more clear and explicit, the present application is further described in detail below in combination with specific embodiments.

[0019] REFERENCE Figure 1The application provides a multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method, which is realized by five core function modules: a multi-source heterogeneous operation parameter acquisition system, a multi-parameter space-time feature fusion and representation module, a physical information constrained load prediction model, a model prediction control based real-time energy-saving optimization unit, and an online learning and model self-adaptive updating mechanism. Data interaction and instruction cooperation are performed between the modules through an edge computing gateway to form a complete "perception-prediction-decision-execution-feedback" closed-loop control system.

[0020] The method comprises the following steps: Step S1, multi-source heterogeneous data acquisition and preprocessing: A multi-source heterogeneous operation parameter acquisition system is constructed, various sensors deployed at key nodes of the air suspension heat pump system are used to acquire at least eight types of operation parameters in real time, including environmental temperature and humidity, building indoor temperature, air suspension compressor speed, refrigerant flow and pressure, evaporator and condenser inlet and outlet temperature, and historical hourly load data, the sampling frequency is not less than 1 time per minute, and the original data acquired is subjected to signal filtering and outlier elimination preprocessing.

[0021] Specifically, the sensor network comprises high-precision temperature and humidity sensors, pressure sensors and turbine flow meters, wherein the temperature measurement accuracy of the temperature and humidity sensor is ±0.2°C, and the humidity measurement accuracy is ±2%RH; the pressure sensor has a range of 0-3MPa and an accuracy of ±0.5% full scale; the measurement accuracy of the turbine flow meter is ±1% reading.

[0022] All the sensors are installed at standardized interface positions, for example, the environmental temperature and humidity sensor is placed at a shaded ventilation position of the building outer wall, the indoor temperature sensor is distributed at a height of 1.5 meters from the ground in the main functional room, the compressor speed signal is directly read through the built-in encoder, the refrigerant high-pressure side and low-pressure side pressures are acquired by the pressure transmitters installed on the exhaust pipe and the suction pipe, and the inlet and outlet temperatures of the evaporator and the condenser are measured by the PT100 platinum resistance thermometer. The acquired data is transmitted to the edge computing gateway through the industrial Ethernet protocol, and the transmission delay is strictly controlled within 50 milliseconds.

[0023] In the data preprocessing stage, firstly, the original time series is subjected to sliding window median filtering, and the window length is set to 5 sampling points to suppress high-frequency noise; then, three-sigma criterion is used to identify and eliminate outliers, that is, for any parameter sequence, if the sampling value at a certain time deviates from the past 60-minute mean value of the sequence by more than 3 times the standard deviation, it is determined as an abnormal point and is replaced by linear interpolation. The preprocessed data is packaged as structured timestamp data packet, the format is JSON, and the fields include timestamp, parameter type identifier, value, unit and confidence flag, which are called by subsequent modules.

[0024] Step S2, spatio-temporal feature fusion and representation: The pre-processed multi-source parameters are input into a lightweight multi-head spatio-temporal feature extraction module embedded with an attention mechanism. The multi-head spatio-temporal feature extraction module extracts the time series local features of each parameter itself through a time convolution network (TCN) and models the dynamic correlation between parameters through a graph attention network (GAT), and finally generates a fused high-dimensional feature tensor.

[0025] Specifically, a tensor with a shape of (T, N) is input into the module, where T = 60 represents a historical window length of 60 minutes, and N ≥ 8 represents a parameter dimension. First, each parameter channel is independently passed through a one-dimensional time convolution network with a depth of 4, with a convolution kernel size of 3, an expansion factor of 1, 2, 4, and 8 in turn, a ReLU activation function, and an output channel number of 64, thereby capturing local patterns of different time scales, such as short-term trends of temperature changes and daily periodicity of loads; then, the feature matrix output by the TCN is reshaped into a node feature matrix H ∈ R^(N×64) as the input of the graph attention network, where R^(N×64) represents a real matrix with N rows and 64 columns; the adjacency relationship of the graph attention network is not predefined, but is dynamically calculated based on the Pearson correlation coefficient between each pair of parameters within a sliding window, that is, for any two parameters i and j, the correlation coefficient is calculated as follows: ; wherein, , is the de-meaned sequence of parameter i and j in the past 30 minutes, , is the standard deviation of the corresponding sequence; only when | ≥ 0.7, a directed edge is established between nodes i and j, and the weight is initialized to .

[0026] The graph attention mechanism adopts a multi-head structure, with a total of 4 independent attention heads, each with a hidden layer dimension of 64. The attention coefficient is calculated through a learnable weight vector, and the formula is as follows: ; wherein, h is the index of the attention head, is the attention coefficient of node j to node i in the hth head, || is the vector concatenation, is the shared linear transformation matrix, is the normalization exponential function along node j, is the activation function, is the transpose of the weight vector . the learnable weight vector for computing the attention score in the h-th head, with input feature vector of node i and node j.

[0027] Finally, the outputs of the four heads are concatenated and linearly projected to 128 dimensions, forming the fused high-dimensional feature tensor Z ∈ R^(N×128). The entire module is processed by channel pruning and weight sparsification to ensure that the single inference delay on the domestic edge computing device with ARM Cortex-A72 processor is less than 200 milliseconds.

[0028] Step S3, load prediction constrained by physical information: A load prediction model constrained by physical information is constructed, the fused high-dimensional feature tensor is input into the load prediction model, and a load prediction value is obtained; the load prediction model integrates a data-driven long short-term memory network and a differentiable physical calculation graph constructed based on the first law of thermodynamics and system operation constraints, and the prediction value is the weighted sum of the preliminary prediction value of the long short-term memory network and the physical residual term.

[0029] Specifically, a hybrid model integrating an LSTM network and physical prior knowledge is constructed, the feature tensor Z obtained in step S2 is input into the model, which is first compressed into a 128-dimensional vector through global average pooling, and then input into an LSTM network with a two-layer stacked structure. Each layer of the LSTM contains 128 hidden units, and Dropout regularization is introduced between layers with a dropout rate of 0.1 to prevent overfitting. The LSTM network outputs a preliminary load prediction value for the future t time , where X t is the input feature vector at time t; at the same time, the model calculates a differentiable physical residual term in parallel, which explicitly encodes the basic laws of thermodynamics and the operating boundaries of the gas suspension heat pump system: which is composed of two sub-terms, one of which is a residual based on the compressor power characteristic, which calculates the theoretical power consumption deviation at the current speed according to the relationship , where is the compressor power, n is the speed, and the other is a residual based on the evaporator energy balance, which calculates the deviation between the measured heat transfer and the theoretical value according to the relationship , where is the evaporator heat transfer, is the mass flow rate of the refrigerant, is the specific heat capacity of the refrigerant at constant pressure, and are the evaporator inlet and outlet temperatures, respectively; both sub-terms are embedded in the forward calculation graph in a differentiable form, without the need to call an external solver; finally, the calculation formula of the load prediction model is as follows: ; wherein, is the load prediction value at time t, λ is a learnable weight coefficient, the initial value is set to 0.1, and is automatically optimized through back propagation during model training.

[0030] The above hybrid structure ensures that the model not only has strong data-driven fitting capability, but also maintains extrapolation stability under the constraint of physical laws, especially avoiding prediction results that violate the laws of thermodynamics under extreme working conditions.

[0031] Step S4, real-time energy-saving optimization based on model predictive control: The load prediction sequence of multiple future periods is input to the rolling optimization model predictive controller, the model predictive controller takes the minimization of the total system energy consumption as the objective function, and is subject to hard constraints composed of the indoor temperature comfort interval, the compressor speed safety interval and the refrigerant pressure limit, to obtain the optimal control instruction sequence.

[0032] Specifically, referring to Figure 2 The real-time energy-saving optimization based on model predictive control (MPC) is performed, the MPC controller takes 15 minutes as the control period, and optimizes the control instructions in the future 4 periods (i.e. 1 hour), the load prediction sequence of the future 8 periods obtained in step S3 is input to the rolling optimization MPC controller, and the objective function is defined as: ; wherein, is the predicted power consumption of the compressor in the kth period, is the predicted power consumption of the circulating water pump in the kth period, is the predicted power consumption of the fan in the kth period, and the compressor power consumption Instead of directly using the simplified model in the physical residual, it is calculated through the pre-calibrated high-precision polynomial fitting curve, which is in the form of wherein the coefficients a, b, c, d are obtained by least squares fitting from the full working condition test data of the equipment under standard working conditions before leaving the factory, and are stored in the read-only memory of the edge device.

[0033] When optimizing the control instructions for the future 4 periods, the internal model of the MPC controller will simulate and consider the influence of the control actions in these 4 periods on the system states in the 5th to 8th periods, so that at the end of the control time domain, the system state is smooth and stable, rather than suddenly out of control or sharply changing, thereby improving the smoothness and comfort of the entire control process.

[0034] The constraint conditions include three hard constraints: first, the indoor temperature must be maintained within the user-set comfort interval , usually = 24℃, = 26℃; second, compressor speed must be located in its safe operation interval , usually = 3000 rpm, = 12000 rpm; third, refrigerant high-pressure side pressure must not exceed the upper limit of system design , usually = 2.8 MPa.

[0035] The above optimization problem is a nonlinear programming problem, which is solved by using the interior point method, and the convergence tolerance is set to 1x10 -6 , and the maximum number of iterations is 50. After the solution is completed, only the optimal control instructions of the first cycle are executed, including the target speed, the target refrigerant flow and the fan speed, and the remaining instructions are used as the prediction reference. The whole MPC solving process is completed on the edge device, and the time consumption of a single optimization does not exceed 300 milliseconds.

[0036] Step S5, online learning and model adaptive update: After each control cycle ends, the error between the actual load and the predicted load is calculated. When the error exceeds the preset threshold for multiple consecutive cycles, the online update mechanism is triggered, and the latest operation data is used to optimize and adjust part of the parameters of the load prediction model.

[0037] Specifically, after each control cycle ends, the mean absolute percentage error (MAPE) of the actual load and the predicted load is calculated. When the MAPE exceeds 5% for 3 consecutive cycles, the online fine-tuning mechanism is triggered, and the recent 32 groups of operation data are used to update the LSTM output layer and the physical residual weight λ by small batch gradient descent. The batch size is fixed at 32, the learning rate is set to 0.001, the maximum number of iterations is 10, and the L2 norm of the weight change is less than 0.01 as an early stop condition.

[0038] The actual load is obtained by real-time integration of the terminal water flow and the supply and return water temperature difference, and the calculation formula is as follows: ; Wherein, is the actual load, is the mass flow rate of water, is the constant-pressure specific heat capacity of water, is the supply water temperature, is the return water temperature.

[0039] The calculation formula of MAPE is as follows: MAPE= ; Wherein, is the load value predicted by the model, The number of samples is collected.

[0040] When the model performance degradation is detected, the system extracts the complete running data packets of the last 32 15-minute periods from the local cache, including the original sensor data, feature tensors, predicted values, and measured values, to construct a fine-tuning dataset. The fine-tuning process only updates the last layer of the LSTM network's fully connected layer weights and the physical residual weight λ, freezing all other parameters to reduce computational overhead and prevent catastrophic forgetting. The gradient descent uses the SGD optimizer with momentum, with a momentum coefficient of 0.9. After each iteration, the L2 norm of the weight update is checked, and if it is less than 0.01, the process is terminated early. After fine-tuning is complete, the new model takes effect immediately, and an update log is recorded for synchronization with the cloud.

[0041] To further improve the robustness of the system, the method of the present application also designs a lightweight running state evaluation and early warning module, which is independent of the above main prediction process, but is closely coupled. It continuously collects the vibration acceleration signals (sampling rate 10 kHz) at the bearing of the gas suspension compressor and the main motor current signals (sampling rate 5 kHz), extracts frequency domain features through fast Fourier transform, and inputs them into a one-dimensional convolutional neural network (1D-CNN) containing only 3 layers. The CNN has a convolution kernel size of 5, channel numbers of 16, 32, and 32 respectively, and a ReLU activation function. Finally, an abnormal confidence is output through global maximum pooling and a fully connected layer. When the confidence exceeds 90%, the system generates a binary warning flag (0 or 1), and this flag is concatenated as an additional feature dimension to the input tensor in step S2, so as to dynamically correct potential deviations caused by bearing wear or instability during load prediction.

[0042] At the engineering deployment level, the present application adopts a collaborative architecture of edge node-cloud platform. The edge node is responsible for the real-time execution of steps S1 to S4, and all models are converted to TensorRT engine format before deployment and quantized to 16-bit floating point, so that the memory occupancy of the LSTM+GAT hybrid model is less than 50MB, significantly reducing the computational and storage overheads, meeting the resource limitations of mainstream domestic edge computing hardware (such as Huawei Atlas500 and Cambrian MLU220), ensuring efficient and stable operation on the corresponding devices, and the calculation delay is less than 1 second. The cloud platform is responsible for offline model training, historical data archiving, and cross-site knowledge migration. The edge and the cloud communicate through 4G or 5G network encryption, use TLS1.3 protocol, and control the daily upload data volume within 100MB, mainly including daily performance indicators, event logs triggering online updates, and monthly model snapshots.

[0043] It should be noted that the initial training of the load prediction model of the present application needs to cover at least two complete refrigeration and heating seasons, the total amount of data is not less than 100,000 15-minute granularity samples, the loss function is mean square error (MSE) plus L2 regularization term (coefficient is 0.001), the optimizer uses Adam, the initial learning rate is 0.01, and the learning rate decay strategy is adopted, and the original 0.9 times is decayed every 10 epochs.

[0044] The above shows and describes the basic principles and main features of the present application and the advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application, and various changes and improvements can be made to the present application without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method, characterized in that, The method comprises the following steps: Step S1, multi-source heterogeneous data acquisition and preprocessing: Through the sensor network deployed in the gas suspension heat pump system, environmental parameters, equipment operating parameters and historical load data are collected in real time, and the original data are preprocessed by filtering and wild value elimination; Step S2, spatio-temporal feature fusion and representation: The preprocessed multi-source parameters are input into a multi-head spatio-temporal feature extraction module embedded with an attention mechanism, the multi-head spatio-temporal feature extraction module extracts the local features of the time series of each parameter through a time convolution network, and models the dynamic correlation between the parameters through a graph attention network, and finally generates a fused high-dimensional feature tensor; Step S3, load prediction with physical information constraint: The high-dimensional feature tensor is input into a load prediction model to obtain a load prediction value; the load prediction model integrates a data-driven long short-term memory network and a differentiable physical calculation graph constructed based on the first law of thermodynamics and system operation constraints, and the prediction value is the weighted sum of the preliminary prediction value of the long short-term memory network and the physical residual term; Step S4, real-time energy-saving optimization based on model predictive control: The load prediction sequence of multiple future periods is input into a rolling optimization model predictive controller, the model predictive controller takes the minimization of the total energy consumption of the system as the objective function, and is limited by the hard constraints composed of the indoor temperature comfort interval, the compressor speed safety interval and the refrigerant pressure limit value, and the optimal control instruction sequence is obtained; Step S5, online learning and model adaptive updating: After each control period, the error between the actual load and the predicted load is calculated; when the error exceeds a preset threshold for consecutive multiple periods, an online updating mechanism is triggered, and the latest operation data is used to optimize and adjust part of the parameters of the load prediction model.

2. The multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method according to claim 1, characterized in that, The sensor network in step S1 includes high-precision temperature and humidity sensors, pressure sensors and turbine flow meters, wherein the temperature measurement accuracy of the temperature and humidity sensor is ±0.2℃, the humidity measurement accuracy is ±2%RH, the pressure sensor has a range of 0-3MPa and an accuracy of ±0.5% full scale, and the turbine flow meter has an accuracy of ±1% reading. All sensor data is transmitted to an edge computing gateway through an industrial Ethernet, and the transmission delay is less than 50 milliseconds.

3. The multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method according to claim 1, characterized in that, The multi-head spatio-temporal feature extraction module in step S2 includes 4 independent attention heads, each head has a hidden layer dimension of 64, the convolution kernel size of the time convolution network is 3, and the adjacency matrix of the graph attention network is dynamically constructed according to the Pearson correlation coefficient between the parameters, and the correlation coefficient threshold is set to 0.

7.

4. The multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method according to claim 1, characterized in that, In step S3, the long short-term memory network has 2 layers, each layer has 128 hidden units, and the loss rate is set to 0.1; the physical calculation graph explicitly encodes the cubic relationship between compressor power and speed, and the product relationship between evaporator heat transfer and refrigerant flow and temperature difference, and the weight coefficient of the physical residual term is learned through back propagation.

5. The multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method according to claim 1, characterized in that, The objective function of the model predictive controller in step S4 is expressed as the total energy consumption equal to the compressor power consumption plus the circulating water pump power consumption plus the fan power consumption, wherein the compressor power consumption is calculated by the power characteristic curve fitting formula, the solver uses the interior point method, the control period is 15 minutes, the prediction time domain is 8 cycles, and the control time domain is 4 cycles.

6. The multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method according to claim 1, characterized in that, The online updating mechanism in step S5 adopts the small batch gradient descent algorithm, the batch size is fixed at 32, the learning rate is set to 0.001, the upper limit of the number of iterations for each update is 10 times, and the L2 norm of the model weight change is less than 0.01 as an early stopping condition.

7. The multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method according to claim 1, characterized in that, It also includes a running state evaluation and early warning mechanism, which is based on real-time collected vibration data and current harmonic analysis, uses a lightweight one-dimensional convolutional neural network to identify early abnormal states of aerodynamic bearings, when the abnormal confidence exceeds 90%, triggers an early warning and inputs the early warning information as a feature to the load prediction model to correct the prediction bias.

8. The multi-parameter fusion air suspension heat pump AI load forecasting energy-saving control method according to claim 1, characterized in that, The training data of the load prediction model covers at least two complete refrigeration and heating seasons, and the total amount of data is not less than 100,000, the mean square error is used as the loss function during training, and the L2 regularization term is added, the regularization coefficient is 0.001, the Adam optimizer is used for training, and the initial learning rate is 0.

01.

9. The multi-parameter fusion air suspension heat pump AI load prediction energy-saving control method according to claim 1, characterized in that, The method is deployed in a hybrid architecture composed of edge computing nodes and cloud data centers, the edge computing nodes are responsible for real-time data collection, feature fusion, prediction and control instruction generation of steps S1 to S4, the calculation delay is less than 1 second; the cloud data center is responsible for model training and historical data management of step S5, the edge and cloud perform encrypted data synchronization through 4G or 5G network, and the daily average data traffic is not more than 100MB.

10. The multi-parameter fusion air suspension heat pump AI load forecasting energy-saving control method according to claim 1, characterized in that, All models involved in the method are deployed in the format of quantization to 16-bit floating point numbers.

Images