Energy system performance prediction method based on pre-training model
By adopting a pre-trained model-based method in energy system modeling, using template matching and pre-training of neural network models, the problems of poor interpretability and insufficient generalization ability in the existing technology are solved, and higher modeling accuracy and interpretability are achieved.
Patent Information
- Application Number
- CN202411754768.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-03
- Publication Date
- 2025-05-13
AI Technical Summary
The prior art relies on data-driven methods in energy system modeling, lacks system mechanism constraints, resulting in poor interpretability of the model, insufficient generalization ability, and highly dependent on training data for training effects.
The energy system performance prediction method based on the pre-trained model is adopted, and the mechanism expression equation of the system template is obtained through template matching, a neural network model with mechanism constraints is constructed, and the performance prediction model of the system to be tested is finally constructed through fine-tuning.
It improves the interpretability, accuracy and generalization capabilities of the model, and can effectively model energy systems with fewer measurement points and less data, reducing the difficulty of modeling.
Smart Images
Figure CN119988899A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a technology in the field of energy system management, and specifically to an energy system performance prediction method based on a pre-training model. Background Art
[0002] The current energy system has the characteristics of numerous devices and large differences in physical properties, which makes the mechanism-based integrated energy system modeling method computationally complex and inefficient. However, traditional neural network modeling requires a large amount of data as support, which is not effective for newly built or data-lacking energy systems. Summary of the invention
[0003] In view of the fact that the prior art relies on data-driven modeling without considering system mechanism constraints, resulting in the lack of interpretability and generalization ability of the model and the defect that the training effect is highly dependent on training data, the present invention proposes an energy system performance prediction method based on a pre-trained model, which does not require solving complex mechanism equations. While reducing the modeling difficulty for complex energy system models, the interpretability and accuracy of the model are significantly improved.
[0004] The present invention is achieved through the following technical solutions:
[0005] The present invention relates to a method for predicting the performance of an energy system based on a pre-trained model. The method comprises collecting input energy and output energy data of a system to be tested and generating a database after pre-processing. Template matching is performed according to the energy input and output types of the system to be tested, a mechanism expression equation is obtained according to the structure of the matched system template, and a neural network model corresponding to the system template is constructed. The neural network model with mechanism constraints is pre-trained based on the data corresponding to the system template to obtain a pre-trained model. By fine-tuning the pre-trained model, a performance prediction model of the system to be tested is constructed to perform real-time performance prediction.
[0006] The neural network model is constructed in the following way: the mechanism model of the energy system is expanded into the form of a state-space equation, and it is discretized using forward Euler to obtain a fully controllable and observable energy system model expression, and an Elman neural network structure is constructed based on the state-space equation form, and finally an interpretable fast Elman neural network (interpretable shortcut Elman network, Shortcut-ENN) model is obtained, which adopts but is not limited to the technical implementation recorded by Zhou D et al. in "Dynamic simulation of natural gas pipeline network based on interpretable machine learning model" ([J]. Energy, 2022, 253: 124068).
[0007] The template matching means that when there is a unique match between the import and export energy types, the template selection is completed; otherwise, the system data features are extracted through the encoder and compared with the features of the system template with the same input and output form to match the best template.
[0008] The encoder and decoder refer to: the encoder outputs low-dimensional feature data for characterizing the system after dimensionality reduction based on the input system original data, and the decoder restores the original data features of the system based on the low-dimensional feature data, both of which are composed of a multi-layer LSTM network.
[0009] The fine-tuning mentioned above refers to: structurally replicating the pre-trained model, initializing the output layer of the model neural network, and collecting input and output energy data of the system to be tested to train some parameters of the model. Technical Effects
[0010] The present invention is based on the energy system template matching mechanism, establishes a data-driven model of the fusion mechanism through the template mechanism equation, and pre-trains the data-driven model using the template corresponding data set. Compared with the prior art, the present invention improves the interpretability, accuracy and generalization ability of the model, and can model energy systems with few measurement points and little data. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] Figure 1 It is a flow chart of the present invention;
[0012] Figure 2 It is the principle diagram of the present invention;
[0013] Figure 3 is a schematic diagram of an embodiment;
[0014] Figure 4 It is a schematic diagram of encoding and decoding feature extraction in an embodiment;
[0015] Figure 5 It is a schematic diagram of the energy system template structure corresponding to the embodiment;
[0016] Figure 6 It is a daily load fluctuation diagram of the embodiment;
[0017] Figure 7 This is a diagram showing the predicted effects of the embodiment. DETAILED DESCRIPTION
[0018] like Figure 1 As shown, this embodiment relates to an energy system performance prediction method based on a pre-trained model, including:
[0019] Step 1: Data collection and preprocessing: Figure 2 As shown, first, the energy system characteristic monitoring data of no less than 168 hours is read cyclically by the on-site data acquisition equipment and stored in the real-time database. The characteristic monitoring data refers to the import and export energy data of the energy system, which may include time series data of various energy forms such as cold, heat, electricity, and gas, and the data is cleaned and supplemented, and finally normalized.
[0020] Step 2: Identify the energy system type: Figure 2 As shown, the energy system is first identified based on its import and export type. If it can uniquely match the energy system template in the template library, the identification is complete. If there are multiple templates with the same import and export energy type as the system to be tested, feature extraction is performed through an encoder based on a recurrent neural network. By comparing the similarity between the feature vector of the system to be tested and the feature vector of the template in the template library, the template feature vector with the smallest Euclidean distance to the feature vector of the system to be tested is found to obtain the best matching template.
[0021] like Figure 3 As shown in the figure, the energy system of this embodiment includes: more energy conversion and storage links. Different energy element configurations and combinations will affect the external output characteristics of the energy system. In the figure: P and L are the input power and output power of the energy conversion device respectively. The energy system inputs energy carrier α and converts it into β, specifically including: L β =c αβ P α , Wherein: C is the coupling matrix of the energy conversion device, which describes the mapping relationship from input power to output power. It can be continuously transformed into various linear transformations, and can also generate various nonlinear relationships.
[0022] like Figure 4 As shown, the encoder and decoder are both composed of a multi-layer LSTM network, wherein: the encoder converts the high-dimensional time series data (X1, X2, ..., X1) from the energy system inton ) is reduced to vector V, and the decoder restores vector V to high-dimensional time series data (X′1, X′2, ..., X′ n ), when the loss function in the model reaches the minimum, the system eigenvector V can be obtained.
[0023] Step 3: Based on the energy system identification result, read the structural data of the corresponding template from the template library.
[0024] like Figure 5 As shown, the template can obtain its mechanism equation according to its internal component structure and connection relationship.
[0025] The template includes: transformer, distributed power generation, gas turbine, waste heat boiler, gas boiler, heat exchange equipment, power storage equipment, cold storage unit, heat pump, electric refrigeration unit, absorption refrigeration unit and heat storage unit, wherein: the transformer input end is connected to the external power grid capacity input, the output end is connected to the power distribution hub, the distributed power generation input end is usually renewable energy such as wind and solar, the output end is connected to the power distribution hub, the gas turbine input end is connected to the external natural gas grid, the output end is connected to the power distribution hub, the waste heat boiler input end is connected to the gas turbine, the output end is connected to the heat energy distribution hub, the gas boiler input end is connected to the external natural gas grid, the output end It is connected to the heat energy distribution hub, the input end of the heat exchanger is connected to the external heating network, and the output end is connected to the heat energy distribution hub, the input end of the power storage device is connected to the electric energy distribution hub, and the output end is connected to the power load, the input end of the cold storage unit is connected to the cold energy distribution hub, and the output end is connected to the cold load, the input end of the heat pump is connected to the electric energy distribution hub, and the output end is connected to the heat energy distribution hub, the input end of the electric refrigeration unit is connected to the electric energy distribution hub, and the output end is connected to the condensation distribution hub, the input end of the absorption refrigeration unit is connected to the heat energy distribution hub, and the output end is connected to the cold energy distribution hub, the input end of the heat storage unit is connected to the heat energy distribution hub, and the output end is connected to the heat load.
[0026] The energy conversion relationship model of the gas turbine includes: GT =G ng ·H ng ·η E，GT , QH GT =G ng ·H ng ·η HE,GT , where: E GT is the power generation of the gas turbine; QH GT G is the thermal power output of the gas turbine; ng is the natural gas consumption in hours; H ng is the calorific value of natural gas; η E，GT and η HE,GT They represent the power generation efficiency and heat generation efficiency of the gas turbine respectively.
[0027] The power generation efficiency and heat generation efficiency of the gas turbine are both related to the actual load size during the operation of the equipment. The load rate is defined as the ratio of the actual output power to the rated output power. The variable operating condition characteristic model of the gas turbine is: Where: f G T is the load factor of the gas turbine; Represents the rated power generation efficiency of the gas turbine; represents the rated heat production efficiency of the gas turbine; a1, b1, c1 and d1 are the power generation efficiency coefficients of the gas turbine; a2, b2 and c2 are the heat production efficiency coefficients of the gas turbine.
[0028] The energy conversion relationship model of the gas boiler includes: QH Bo =G ng ·H ng ·η QH,Bo , of which: QH Bo is the heating power of the gas boiler; G ng is the natural gas consumption; H ng is the calorific value of natural gas; η QH,Bo The purpose is to achieve the overall efficiency of boiler heat production, which is determined by the heat loss of each part.
[0029] The energy conversion relationship model of the absorption refrigerator includes: QC AC =COP QC,AC ·QH AC , where: QC AC is the output cooling power of the absorption chiller; QH AC is the thermal power input to the absorption chiller; COP QC,AC It is its refrigeration coefficient, which reflects the refrigeration performance of the refrigerator.
[0030] The refrigeration coefficient COP of the absorption refrigerator QC,AC It is mainly related to the load rate, and its variable operating condition model is: in: is the rated refrigeration factor of the absorption chiller; f AC is the load rate of the absorption chiller; a5, b5, c5 and d5 are the influencing factors of the refrigeration coefficient of the absorption chiller.
[0031] The energy conversion relationship model of the compression refrigerator includes: QC CR =COP QC,CR ·E CR , where: QC CR is the cooling power output of the compression refrigerator; E CR The power consumption of the compression refrigerator; COP QC,CRIt is its cooling efficiency.
[0032] The refrigeration coefficient of the compression refrigerator is also closely related to its load rate. According to the empirical formula, the variable operating condition model of the compression refrigerator is: in: is the rated refrigeration factor of the compression refrigerator; f CR is the load rate of the compression refrigerator, a6, b6 and c6 are the influencing factors of the refrigeration coefficient of the compression refrigerator.
[0033] The energy conversion relationship model of photovoltaic power generation includes: Where: E PV is the photovoltaic output power; H s is the actual light radiation intensity; T c is the actual temperature of the photovoltaic cell; T r is the reference temperature of the photovoltaic cell, usually 25°C; the subscript stc is the standard test condition, E stc , H stc They are the maximum photovoltaic output power and light radiation intensity under the corresponding standard test conditions; k is the power temperature coefficient of the photovoltaic panel.
[0034] In actual operation, the power generation of solar panels and the light radiation density are basically linearly related, and a photovoltaic power plant is jointly produced by multiple solar photovoltaic panels, so the total power generation of the photovoltaic power plant is E PV =η PV ·H s ·A p N1 / H stc , where: A p is the area of a single photovoltaic panel; N1 is the total number of photovoltaic panels; η PV It is the conversion efficiency coefficient of solar photovoltaic power generation.
[0035] The energy conversion relationship model of wind power generation includes: Where: E WPG is the output power of the wind turbine; E r is the rated power of the wind turbine; v ci and v co are the cut-in wind speed and the cut-out wind speed respectively; v r is the rated wind speed of the wind turbine; v is the actual wind speed.
[0036] The energy conversion relationship model of the battery includes: the process of storing electric energy The process of releasing electrical energy Where: SOC(t) is the remaining energy storage capacity of the battery at the end of the tth moment; SOC(t-1) is the remaining energy storage capacity of the battery at the end of the (t-1)th moment; δ SB is the battery’s own power consumption rate; is the power stored in the battery at the tth moment; is the power released by the battery at the tth moment; The energy storage efficiency of the battery; The energy release efficiency of the battery; is the rated capacity of the battery.
[0037] Taking into account the actual conditions of battery capacity and the rate of storing and releasing electric energy, the constraints satisfied by the parameters in the battery energy conversion relationship model include: in: is the maximum electrical energy storage capacity of the battery; The maximum power stored in the battery; It is the maximum electrical output power of the battery.
[0038] The energy conversion relationship model of the heat storage tank includes: the heat storage process Heat release process Among them: QH hst (t) is the heat energy stored in the heat storage tank at time t, is the heat absorption capacity of the heat storage tank at time t, is the heat release capacity of the heat storage tank at time t, μ1 is the heat energy loss coefficient of the heat storage tank itself, and They represent the heat absorption efficiency and heat release efficiency of the heat storage tank. In this system, the heat storage tank is set to complete only one of the functions of heat absorption or heat release at the same time.
[0039] Taking into account the actual conditions of the heat storage tank capacity and the heat absorption and release rates, the constraints satisfied by the parameters in the energy conversion relationship model of the heat storage tank include: in: The upper and lower bounds and QH hst The upper limit constraints of (t) are all determined by the characteristics of the equipment.
[0040] The energy conversion relationship model of the cold storage box includes: the cold energy storage process Release of cold energy Among them: QC cst (t) is the cold energy stored in the cold storage box at time t, is the cooling power stored in the cold storage box at time t, is the cooling power released by the cold storage box at time t, μ2 is the cooling energy loss coefficient of the cold storage box itself, and They represent the cold energy input conversion efficiency and cold energy output conversion efficiency of the cold storage box respectively. Considering the actual situation of the cold storage tank capacity and the storage and release rate of cold energy, the constraints satisfied by the parameters of the cold storage tank energy conversion relationship model include: in: Upper and lower bound constraints and QC cst The upper limit constraints of (t) are all determined by the characteristics of the equipment.
[0041] Step 4: Construct a neural network model of the template: Figure 2 As shown, through the mechanism equation of the model, the standard state space equation of the component consistent with the feedforward characteristics of the neural network is derived, and the neural network model is constructed according to the mechanism equation and parameter transfer relationship to achieve efficient training and information extraction of the energy system model, and finally establish a data-driven model that integrates mechanism knowledge.
[0042] Step 5: Construct a pre-trained model: Figure 2 As shown in the figure, the pre-training data set corresponding to the template is selected from the pre-training data set to train the data-driven model integrating mechanism knowledge. The mean square error (MSE) is used as the loss function indicator to measure the model training effect. The loss function is calculated with the real data P′, and the calculation method is as follows: The model parameters are adjusted according to the loss function until the training results converge and the loss function reaches a stable minimum. The data-driven model of the trained template is structurally replicated and fine-tuned, and the output layer parameters are initialized to obtain a pre-trained model.
[0043] The fine-tuning is specifically as follows: copy all the structures and weight parameters of the data-driven model of the template, and restore the parameters of the output layer to the initial values when there is no difference between the output dimensions and variable types of the template and the system to be tested. When there is a difference between the output dimensions or variable types of the template and the system to be tested, delete the structure and parameters of the output layer of the original template, and construct a new output layer based on the output dimensions of the previous layer of the output layer of the model and the output dimensions of the system to be tested. The input dimensions and output dimensions of the new output layer are respectively consistent with the output dimensions of the previous layer and the output dimensions of the system to be tested. Finally, a new model is obtained, which is the pre-trained model.
[0044] Step 6: Establish the system model to be tested: Figure 2As shown in the figure, the pre-trained model is trained based on the real input and output energy data of the system to be tested, the real load data L is used as the model input, and the energy input prediction value output by the model is The loss function is calculated with the real energy input data P, and the calculation method is as follows: The model is self-supervised learning using MSE as the loss function until the loss function on the test set is stable and reaches the minimum, at which time the performance prediction model of the system to be tested is obtained.
[0045] Step 7: Input the boundary condition parameters of the model to be tested under the required simulation conditions into the performance prediction model, and the model will calculate the simulation results of the target energy system.
[0046] After specific practical experiments, the specific settings are as follows: based on a small amount of data from the target system for one week, the data time resolution is 10 minutes, and the historical 60-minute long export parameter sequence is used to predict the import energy input parameters 10 minutes later. The training process uses a learning rate of 0.0014. Figure 6 The electrical load, heat load, and cooling load shown are boundary conditions. Figure 7 As shown in the figure, the power, natural gas, and heat input in the simulation results are very consistent with the real data, and the model accuracy is high. The MSE of the simulation results was calculated, and the MSE of the natural gas, heat, and power input prediction results of this method were 0.002, 0.0006, and 0.0021, respectively, which proves the accuracy and effectiveness of this method in energy system model prediction.
[0047] Compared with the existing technology, this method is based on the energy system template matching mechanism to achieve pre-training of interpretable neural networks with mechanism implantation, which improves the interpretability, accuracy and generalization ability of the model.
[0048] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.
Claims
1. A method for predicting energy system performance based on a pre-trained model, characterized in that: A database is generated by collecting input energy and output energy data of the system to be tested and preprocessing them; template matching is performed according to the energy input and output types of the system to be tested, and the mechanism expression equation is obtained according to the structure of the matched system template, and a neural network model corresponding to the system template is constructed, and the neural network model with mechanism constraints is pre-trained based on the data corresponding to the system template to obtain a pre-trained model, and by fine-tuning the pre-trained model, a performance prediction model of the system to be tested is constructed to perform real-time performance prediction.
2. The energy system performance prediction method based on the pre-training model according to claim 1 is characterized in that: The neural network model is constructed in the following way: the mechanism model of the energy system is expanded into the form of state space equations, and it is discretized using forward Euler to obtain a fully controllable and observable energy system model expression, and an Elman neural network structure is constructed based on the state space equation form, and finally an interpretable fast Elman neural network model is obtained.
3. The energy system performance prediction method based on the pre-training model according to claim 1 is characterized in that: The template matching means that when there is a unique match between the import and export energy types, the template selection is completed; otherwise, the system data features are extracted through the encoder and compared with the features of the system template with the same input and output form to match the best template.
4. The energy system performance prediction method based on the pre-training model according to claim 1 is characterized in that: The fine-tuning mentioned above refers to: structurally replicating the pre-trained model, initializing the output layer of the model neural network, and collecting input and output energy data of the system to be tested to train some parameters of the model.
5. The energy system performance prediction method based on the pre-training model according to any one of claims 1 to 4 is characterized in that: include: Step 1, data collection and preprocessing: First, the energy system’s characteristic monitoring data of no less than 168 hours is cyclically read through the on-site data collection equipment and stored in the real-time database. The characteristic monitoring data refers to the import and export energy data of the energy system, which may include time series data of various energy forms such as cold, heat, electricity, and gas. The data is cleaned and supplemented, and finally normalized; Step 2, Identify the energy system type: First, identify the energy system import and export type. If it can be uniquely matched with the energy system template in the template library, the identification is completed. If there are multiple templates with the same import and export energy type as the system to be tested, feature extraction is performed through an encoder based on a recurrent neural network. By comparing the similarity between the feature vector of the system to be tested and the feature vector of the template in the template library, the template feature vector with the smallest Euclidean distance to the feature vector of the system to be tested is found to obtain the best matching template; The energy system includes: more energy conversion and storage links. Different energy element configurations and combinations will affect the external output characteristics of the energy system. In the figure: P and L are the input power and output power of the energy conversion device respectively. The energy system input energy carrier α is converted into β, specifically including: L β =c αβ P α , Where: C is the coupling matrix of the energy conversion device, which describes the mapping relationship from input power to output power. It continuously transforms various linear transformations and also produces various nonlinear relationships; Step 3: Based on the energy system identification result, read the structural data of the corresponding template from the template library; The template includes: transformer, distributed power generation, gas turbine, waste heat boiler, gas boiler, heat exchange equipment, power storage equipment, cold storage unit, heat pump, electric refrigeration unit, absorption refrigeration unit and heat storage unit, wherein: the transformer input end is connected to the external power grid capacity input, the output end is connected to the power distribution hub, the distributed power generation input end is usually renewable energy such as wind and solar, the output end is connected to the power distribution hub, the gas turbine input end is connected to the external natural gas grid, the output end is connected to the power distribution hub, the waste heat boiler input end is connected to the gas turbine, the output end is connected to the heat energy distribution hub, the gas boiler input end is connected to the external natural gas grid, the output end connected to the heat energy distribution hub, the heat exchange equipment input end is connected to the external heating network, and the output end is connected to the heat energy distribution hub, the power storage equipment input end is connected to the electric energy distribution hub, and the output end is connected to the power load, the cold storage unit input end is connected to the cold energy distribution hub, and the output end is connected to the cold load, the heat pump input end is connected to the electric energy distribution hub, and the output end is connected to the heat energy distribution hub, the electric refrigeration unit input end is connected to the electric energy distribution hub, and the output end is connected to the condensation distribution hub, the absorption refrigeration unit input end is connected to the heat energy distribution hub, and the output end is connected to the cold energy distribution hub, the heat storage unit input end is connected to the heat energy distribution hub, and the output end is connected to the heat load; Step 4: Construct a neural network model of the template: Through the mechanism equation of the model, derive the standard state space equation of the component consistent with the feedforward characteristics of the neural network, build a neural network model based on the mechanism equation and parameter transfer relationship, realize efficient training and information extraction of the energy system model, and finally establish a data-driven model integrating mechanism knowledge; Step 5: Construct a pre-trained model: Select the pre-training data set corresponding to the template from the pre-training data set to train the data-driven model integrating the mechanism knowledge. Use the mean square error (MSE) as the loss function indicator to measure the model training effect. Calculate the loss function with the real data P', the calculation method is as follows: Adjust the model parameters according to the loss function until the training results converge and the loss function reaches a stable minimum, perform structural replication and fine-tuning on the data-driven model of the trained template, and initialize the output layer parameters to obtain a pre-trained model; Step 6: Establish the model of the system to be tested: Train the pre-trained model based on the real input and output energy data of the system to be tested, use the real load data L as the model input, and use the energy input prediction value output by the model The loss function is calculated with the real energy input data P, and the calculation method is as follows: The model is self-supervised learning with MSE as the loss function until the loss function on the test set is stable and reaches the minimum, at which time the performance prediction model of the system to be tested is obtained; Step 7: Input the boundary condition parameters of the model to be tested under the required simulation conditions into the performance prediction model, and the model will calculate the simulation results of the target energy system.
6. The energy system performance prediction method based on the pre-training model according to claim 5 is characterized in that: The energy conversion relationship model of the gas turbine includes: GT =G ng ·H ng ·η E,GT , QH GT =G ng ·H ng ·η HE,GT , where: E GT is the power generation of the gas turbine; QH GT G is the thermal power output of the gas turbine; ng is the natural gas consumption in hours; H ng is the calorific value of natural gas; η E,GT and η HE,GT They represent the power generation efficiency and heat generation efficiency of the gas turbine respectively; The power generation efficiency and heat generation efficiency of the gas turbine are both related to the actual load size during the operation of the equipment. The load rate is defined as the ratio of the actual output power to the rated output power. The variable operating condition characteristic model of the gas turbine is: Where: f GT is the load factor of the gas turbine; Represents the rated power generation efficiency of the gas turbine; represents the rated heat production efficiency of the gas turbine; a1, b1, c1 and d1 are the power generation efficiency coefficients of the gas turbine; a2, b2 and c2 are the heat production efficiency coefficients of the gas turbine; The energy conversion relationship model of the gas boiler includes: QH Bo =G ng ·H ng ·η QH,Bo , of which: QH Bo is the heating power of the gas boiler; G ng is the natural gas consumption; H ng is the calorific value of natural gas; η QH,Bo The purpose is to determine the total efficiency of the boiler in producing heat, which is determined by the heat losses of each part; The energy conversion relationship model of the absorption refrigerator includes: QC AC =COP QC,AC ·QH AC , where: QC AC is the output cooling power of the absorption chiller; QH AC is the thermal power input to the absorption chiller; COP QC,AC It is its refrigeration coefficient, which reflects the refrigeration performance of the refrigerator; The refrigeration coefficient COP of the absorption refrigerator QC,AC It is mainly related to the load rate, and its variable operating condition model is: in: is the rated refrigeration factor of the absorption chiller; f AC is the load rate of the absorption chiller; a5, b5, c5 and d5 are the factors affecting the cooling coefficient of the absorption chiller; The energy conversion relationship model of the compression refrigerator includes: QC CR =COP QC,CR ·E CR , where: QC CR is the cooling power output of the compression refrigerator; E CR The power consumption of the compression refrigerator; COP QC,CR is its cooling efficiency; The refrigeration coefficient of the compression refrigerator is also closely related to its load rate. According to the empirical formula, the variable operating condition model of the compression refrigerator is: in: is the rated refrigeration factor of the compression refrigerator; f CR is the load rate of the compression refrigerator, a6, b6 and c6 are the influencing factors of the refrigeration coefficient of the compression refrigerator; The energy conversion relationship model of photovoltaic power generation includes: Where: E PV is the photovoltaic output power; H s is the actual light radiation intensity; T c is the actual temperature of the photovoltaic cell; T r is the reference temperature of the photovoltaic cell, usually 25°C; the subscript stc is the standard test condition, E stc , H stc are the maximum photovoltaic output power and light radiation intensity under the corresponding standard test conditions; k is the power temperature coefficient of the photovoltaic panel; The energy conversion relationship model of wind power generation includes: Where: E WPG is the output power of the wind turbine; E r is the rated power of the wind turbine; v ci and v co are the cut-in wind speed and the cut-out wind speed respectively; v r is the rated wind speed of the wind turbine; v is the actual wind speed; The energy conversion relationship model of the battery includes: the process of storing electric energy The process of releasing electrical energy Where: SOC(t) is the remaining energy storage capacity of the battery at the end of the tth moment; SOC(t-1) is the remaining energy storage capacity of the battery at the end of the (t-1)th moment; δ SB is the battery’s own power consumption rate; is the power stored in the battery at the tth moment; is the power released by the battery at the tth moment; The energy storage efficiency of the battery; The energy release efficiency of the battery; is the rated capacity of the battery; The energy conversion relationship model of the heat storage tank includes: the heat storage process Heat release process Among them: QH hst (t) is the heat energy stored in the heat storage tank at time t, is the heat absorption capacity of the heat storage tank at time t, is the heat release capacity of the heat storage tank at time t, μ1 is the heat energy loss coefficient of the heat storage tank itself, and They represent the heat absorption efficiency and heat release efficiency of the heat storage tank respectively. In this system, the heat storage tank is set to complete only one of the functions of heat absorption or heat release at the same time; The energy conversion relationship model of the cold storage box includes: the cold energy storage process Release of cold energy Among them: QC cst (t) is the cold energy stored in the cold storage box at time t, is the cooling power stored in the cold storage box at time t, is the cooling power released by the cold storage box at time t, μ2 is the cooling energy loss coefficient of the cold storage box itself, and They represent the cold energy input conversion efficiency and cold energy output conversion efficiency of the cold storage box respectively.
7. The energy system performance prediction method based on the pre-training model according to claim 5 is characterized in that: The fine-tuning is specifically as follows: copy all structures and weight parameters of the data-driven model of the template; when there is no difference in the output dimension and variable type between the template and the system to be tested, restore the parameters of the output layer to the initial values; when there is a difference in the output dimension or variable type between the template and the system to be tested, delete the structure and parameters of the original template output layer; and construct a new output layer according to the output dimension of the previous layer of the output layer of the model and the output dimension of the system to be tested; the input dimension and output dimension of the new output layer are respectively consistent with the output dimension of the previous layer and the output dimension of the system to be tested, and finally obtain a new model, which is the pre-trained model.