Physically guided and data-driven integrated energy system equipment modeling method and system

Abstract

The application provides a kind of physical guidance and data-driven integrated energy system equipment modeling method and system, convolutional neural network and gated recurrent unit neural network are combined to form hybrid data-driven model;Mechanism model of integrated energy system equipment is constructed, data enhancement of mechanism model is carried out, simulation data is generated, and the hybrid data-driven model is pre-trained using the simulation data;Based on the physical law of the integrated energy system equipment, the physical constraint condition is constructed and modeled into the loss function of the hybrid data-driven model, and the hybrid data-driven model is trained using the actual data to obtain the final modeling result.The application utilizes the synergy of mechanism model and data model in the model construction pre-training process and the training process, realizes the hybrid modeling of integrated energy system equipment, and guarantees the accuracy and precision.

Description

Technical Field

[0001] This invention belongs to the field of integrated energy system modeling technology, and relates to a physical-guided and data-driven integrated energy system equipment modeling method and system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Integrated Energy Systems (IES) combine cyber-physical systems, multi-energy supply technologies, and diverse energy storage technologies such as thermal, electrical, and gas storage. They organically integrate thermal, electrical, and gas networks to achieve multi-energy conversion, storage, and consumption. This is of great significance for improving the efficiency of comprehensive energy utilization, promoting the large-scale development of renewable energy, improving the utilization rate of social infrastructure, and ensuring energy supply security.

[0004] Therefore, modeling key equipment, energy stations, and multi-energy flow at each level of the IES is crucial. Constructing multi-dimensional models can meet the needs of characteristic analysis, planning and design, and operation control. Existing conventional modeling methods include mechanistic modeling and data modeling. The shortcomings of mechanistic models are that for some objects, the mechanistic model consists of high-order differential equations, which are difficult to solve; and for some objects, it is difficult to obtain accurate mathematical expressions, resulting in models that differ significantly from actual equipment. The shortcomings of data models are that they require a large amount of training data and do not consider any physical laws (such as the conservation of energy or mass), which may produce results inconsistent with physical laws. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a physical-guided and data-driven integrated energy system equipment modeling method and system. In both the pre-training and training processes of model building, this invention utilizes the synergy of mechanistic models and data models to achieve hybrid modeling of integrated energy system equipment. This avoids errors caused by the knowledge limitations of mechanistic models and reduces physical inconsistencies that exist in empirical models built through machine learning methods.

[0006] According to some embodiments, the present invention adopts the following technical solution:

[0007] A physical-guided and data-driven integrated energy system equipment modeling method includes the following steps:

[0008] By combining convolutional neural networks and gated recurrent unit neural networks, a hybrid data-driven model is formed.

[0009] A mechanism model of integrated energy system equipment is constructed, data augmentation of the mechanism model is performed, simulation data is generated, and the simulation data is used to pre-train the hybrid data-driven model.

[0010] Based on the physical laws of the integrated energy system equipment, physical constraints are constructed and modeled into the loss function of the hybrid data-driven model. The hybrid data-driven model is then trained using actual data to obtain the final modeling result.

[0011] The above technical solution involves at least two fusions of the mechanistic model and the data model. During pre-training, data from the mechanistic model is used to pre-train the data model. During training, the physical knowledge of the mechanistic model is used to constrain the loss function of the data model. This avoids errors caused by the knowledge limitations of the mechanistic model and reduces physical inconsistencies inherent in empirical models built through machine learning methods. Furthermore, the data model is a CNN-GRU hybrid neural network model, combining the advantages of both to ensure higher accuracy and precision in the final device model, enabling more accurate modeling of device characteristics and performance.

[0012] As an alternative implementation, the hybrid data-driven model includes an input layer, a CNN layer, a GRU layer, a fully connected layer, and an output layer set up in one step. The CNN layer includes several convolutional layers and pooling layers that are alternately stacked, and there is at least one GRU layer.

[0013] As a further step, the convolutional layer is a one-dimensional convolution, the convolution method is selected as Same convolution, and the activation function is selected as ReLU function; the pooling layer uses Valid max pooling operation as the pooling method, and the fully connected layer uses the Sigmoid activation function.

[0014] As an alternative implementation method, when constructing the mechanism model of the integrated energy system equipment, the mechanism models of volumetric flow rate and output torque are established by comprehensively considering the effects of leakage, over- and under-expansion and friction factors, respectively. The speed, intake pressure and / or temperature are used as control variables to establish a mathematical model of volumetric flow rate and / or output torque with respect to the control variables.

[0015] As an alternative implementation, generative adversarial networks (GANs) are used to augment the mechanistic model. The GAN includes a generator and a discriminator. The generator receives random variables and generates data samples. The discriminator determines whether the input samples are real or synthetic. The discriminator also calculates the loss and performs backpropagation to obtain the gradient, thereby updating the parameters.

[0016] As an alternative implementation, when training the hybrid data-driven model, the Adam algorithm is used to iteratively update the weights. The weights and biases of each neuron are continuously updated through momentum and adaptive learning rate, so that the output value of the loss function reaches the optimal value. The loss function uses the mean squared error function.

[0017] As an alternative implementation, the specific process of constructing physical constraints based on the physical laws of the integrated energy system equipment includes constructing physical constraints with energy conservation and / or mass conservation, embedding constraints through a loss function, and adjusting the proportion of constraints in the overall loss function using weighting coefficients.

[0018] A physical-guided and data-driven integrated energy system equipment modeling system, comprising:

[0019] The hybrid data-driven model building module is configured to combine convolutional neural networks and gated recurrent unit neural networks to form a hybrid data-driven model.

[0020] The pre-training fusion module is configured to construct a mechanism model of integrated energy system equipment, perform data augmentation on the mechanism model, generate simulation data, and use the simulation data to pre-train the hybrid data-driven model.

[0021] The training fusion module is configured to construct physical constraints based on the physical laws of the integrated energy system equipment, and model them into the loss function of the hybrid data-driven model. The hybrid data-driven model is then trained using real data to obtain the final modeling result.

[0022] A computer-readable storage medium storing a plurality of instructions adapted for loading by a processor of a terminal device and executing steps in the method.

[0023] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement instructions; the computer-readable storage medium being configured to store a plurality of instructions adapted to be loaded by the processor and executed in accordance with the steps of the method described therein.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] The physics-guided CNN-GRU hybrid neural network modeling method proposed in this invention is simpler and more universal than traditional mechanistic models.

[0026] The hybrid modeling method provided by this invention avoids errors caused by the knowledge limitations of mechanistic models and reduces physical inconsistencies in empirical models built through machine learning. Validating this modeling method using a vortex expander as an example, it can be seen that compared to traditional neural network modeling methods, the model built by this invention has smaller errors and higher modeling accuracy, indicating that the established model can more accurately model the characteristics and performance of the equipment.

[0027] This invention has wide applications and is suitable for the modeling process of key devices at the bottom layer of integrated energy systems. Attached Figure Description

[0028] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0029] Figure 1 This is a technical roadmap of the modeling method proposed in this embodiment;

[0030] Figure 2 This is the CNN-GRU model structure proposed in this embodiment;

[0031] Figure 3 This is a schematic diagram of the working principle of the vortex expander in this embodiment;

[0032] Figure 4 This is a graph showing the change in expansion efficiency with rotational speed under different intake pressures in the model established in this embodiment;

[0033] Figure 5(a) is a structural diagram of the experimental system built in this embodiment;

[0034] Figure 5(b) is a physical distribution diagram of the experimental platform built in this embodiment;

[0035] Figure 6 This is a graph showing the change in root mean square error during the model training process of the hybrid modeling method in this embodiment;

[0036] Figure 7(a) shows the volumetric flow rate model established in this embodiment;

[0037] Figure 7(b) shows the output torque model established in this embodiment;

[0038] Figure 8 This is a radar chart comparing and verifying multiple modeling methods in this embodiment. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] Example 1

[0043] A physics-guided CNN-GRU hybrid neural network modeling method. Firstly, the technical route of the modeling method designed in this invention is as follows: Figure 1 As shown.

[0044] 1. CNN-GRU hybrid neural network model

[0045] The CNN-GRU model structure built in this invention is as follows: Figure 2 As shown, it is mainly divided into an input layer, a CNN layer, a GRU layer, and an output layer.

[0046] (1) Convolutional Neural Network

[0047] CNN is a type of feedforward neural network with a convolutional structure. CNN models reduce the number of parameters and alleviate overfitting by sharing weights. By alternating convolutional and pooling layers with local connections between neurons, CNNs perform higher-level and more abstract processing of the initial data, effectively extracting internal features.

[0048] The internal neural network layers of a CNN model mainly consist of an input layer, a convolutional layer (CONV), a pooling layer (POOL), a fully connected layer (FC), and an output layer. The core structure is the repeated use of convolutional and pooling layers. This structure "flattens" the multidimensional data, reduces the number of weights, effectively reduces the complexity of feature extraction and data reconstruction, and improves the quality of data features.

[0049] (2) Gated neural units

[0050] GRU is an optimization and improvement on the Long Short-Term Memory (LSTM) neural network. It integrates the LSTM forget gate and input gate into a single update gate. Compared with LSTM, GRU reduces training parameters, shortens model training time, and maintains accuracy while more effectively reducing the risk of overfitting.

[0051] The mathematical description is as follows:

[0052] z t =σ(W z ·[h t-1 ,x t ]+b z (1)

[0053] r t =σ(W r ·[h t-1 ,x t ]+b r (2)

[0054] h′ t =tanh(W c ·[r t ×h t-1 ,x t ]+b c (3)

[0055] h t =(Iz t )×h t-1 +z t ×h′ t (4)

[0056] In the formula: x t For the input vector, h t-1 h is the historical state variable from the previous time step. t The state variable z at the current moment t To update the state of the gate, r t To reset the state of the door, h′ t b is the current candidate hidden state; z ,b r ,b c W is the bias value. z W r W c Let I be the weight parameter matrix; I represents the identity matrix; [*] represents vector concatenation; · represents matrix dot product; × represents matrix multiplication; σ(*) represents the sigmoid activation function, and tanh(*) is the activation function, mathematically described as follows:

[0057]

[0058]

[0059] The GRU network uses update gates and reset gates as its core modules, with input variable x. t The state memory variable h from the previous time step t-1 The concatenated matrix, after nonlinear transformation by the sigmoid activation function, is input into the update gate, determining which state variables from the previous time step are incorporated into the current state. The reset gate controls the amount of information written into the candidate hidden state from the previous time step, via Iz... t times h t-1 Store the information from the previous moment, via z t times h′ t Record the information at the current moment and obtain the output yt at the current moment:

[0060] y t =W o h t +b o (7)

[0061] In the formula: W o b is the weight parameter; o This is the bias value.

[0062] (3) Hybrid Neural Network Model

[0063] The network structure is as follows: The CNN network consists of two convolutional layers and two pooling layers. Based on the characteristics of the experimental data, the convolutional layers are designed as one-dimensional convolutions, using Same convolution and ReLU activation. To retain more data fluctuation information, the pooling layers use Valid Max Pooling. The GRU recurrent neural network learns from the extracted feature vectors; a two-layer GRU structure achieves the best modeling effect. Finally, the output of the fully connected layer is inversely normalized to obtain the final output. The fully connected layer uses the Sigmoid activation function.

[0064] Different features have different dimensions and large numerical differences, which is not conducive to the optimization of model parameters. This paper uses the values ​​and minimum values ​​of features in the training set as a benchmark and normalizes the data according to the following formula.

[0065]

[0066] In the formula: x * x represents the normalized data; x represents the original data; x max x min These represent the maximum and minimum values ​​of the data in each dimension of the training set.

[0067] During training of the GRU recurrent neural network, the Adam algorithm is used to iteratively update the weights. By continuously updating the weights and biases of each neuron through momentum and adaptive learning rate, the output value of the loss function is optimized. Adam is an algorithm that can replace the traditional stochastic gradient descent process. This algorithm can iteratively update the weights of the neural network based on the training data, maximizing the output value of the loss function. The model's loss function uses the mean squared error function, i.e.

[0068]

[0069] In the formula: n is the number of samples; y i Actual value; Output values ​​for the model.

[0070] 2. Pre-training

[0071] Data models typically require parameter initialization before training. Poor initialization can cause the model to converge to a local optimum, especially for neural network modeling methods. If weights can be initialized based on the device's mechanistic model, it can improve training efficiency (i.e., reduce training time) and reduce the amount of experimental data required, achieving good performance with fewer data samples.

[0072] (1) Mechanism Model

[0073] Due to the complexity of experimental data, taking the vortex expander in this paper as an example, in addition to directly measurable parameters such as rotational speed, inlet temperature, volumetric flow rate, and output torque, there are many time-varying and distributed parameters, as well as some process states that are difficult to measure, among which leakage, friction, and heat transfer are the main factors. To solve these problems, this embodiment uses simulation data generated by a vortex expander model that comprehensively considers the influencing factors such as leakage, friction, and heat transfer to pre-train a CNN-GRU model.

[0074] Of course, this embodiment is merely an example. In other embodiments, the method provided by this invention can be used to build models of equipment in other integrated energy systems. However, adjustments to the considerations or expressions need to be made based on the physical characteristics of the equipment, which are all things that those skilled in the art can conceive of.

[0075] A novel compressed air energy storage system based on a scroll compressor operates on the principle of storing excess energy as compressed air in a tank using a scroll compressor. When needed, the energy is released by an expander to compensate for power fluctuations in the microgrid. The main components of the scroll expander are a moving scroll and a stationary scroll. The moving scroll is driven by a motor to rotate, while the stationary scroll remains stationary. These two scrolls form several crescent-shaped expansion chambers with a 180° phase difference. The moving scroll rotates under the drive of a prime mover, drawing gas into the chambers from points A and B. During rotation, the gas is compressed layer by layer as the chamber volume decreases, eventually being forced into chamber I. When the gas pressure in chamber I exceeds the exhaust pressure, the exhaust valve opens, and the gas is stored in the storage tank. The working principle of the scroll compressor is as follows: Figure 3 .

[0076] When analyzing the mechanism model of the vortex expander, the mechanism models of volumetric flow rate and output torque are established by comprehensively considering the effects of leakage, over- and under-expansion and friction. The easily measurable speed, intake pressure and temperature are used as control variables, and mathematical models of volumetric flow rate and output torque with respect to these variables are established.

[0077] 1) Volumetric flow rate model

[0078] Inlet volumetric flow rate per unit time of a vortex expander:

[0079]

[0080] In the formula, n is the spindle speed of the expander, in r / min.

[0081] The flow loss caused by leakage under standard conditions is:

[0082]

[0083] In the formula, Se is the effective cross-sectional area of ​​the leakage gap.

[0084] Volumetric flow rate mechanism model:

[0085]

[0086] 2) Output torque model

[0087] The viscous friction torque of the gas is directly proportional to the rotational speed of the vortex expander as follows:

[0088]

[0089] In the formula, Bt is the gas viscosity coefficient, the magnitude of which is related to the gas temperature. The system studied in this paper operates over a relatively small temperature range, so it can be considered a constant.

[0090] The frictional torque is:

[0091]

[0092] In the formula, M1, M2, and M3 correspond to the friction torque at the three locations mentioned above. μ1 is the friction coefficient between the main shaft crank pin and the driven surface, md is the mass of the moving scroll plate, ror is the crank rotation radius, μ2 is the equivalent friction coefficient at the cross slip ring, which can be considered a constant value. Therefore, the friction torque M2 at the contact point between the cross slip ring groove and the side of the cross slip ring key is proportional to the intake pressure. μ3 is the friction coefficient between the moving scroll plate and the stationary scroll plate base plate, which is also considered a constant value. Mt is the overturning moment, and Do are parameters related to the number of spiral coils, the base circle radius, and the spiral tooth wall thickness.

[0093] The volumetric flow rate model and output torque model are used to express the actual output torque of the vortex expander.

[0094]

[0095] After establishing the volumetric flow rate model and the mechanistic model of the output torque, intake pressure, speed, and intake temperature are selected as the input variables of the model, and volumetric flow rate and output torque are the output variables.

[0096] (2) Gans data augmentation

[0097] GANs consist of two parts: a generator and a discriminator. The generator receives random variables and generates data samples, while the discriminator determines whether the input samples are real or synthetic. They work against each other to improve each other's performance. The discriminator performs binary classification tasks, calculating the loss and backpropagating to find the gradient, thereby updating the parameters.

[0098] Because pre-trained neural networks require sufficient data, after establishing the mechanistic model, GANs are used to augment the dataset by performing data augmentation on the established mechanistic model. The GAN generator receives the vortex expander mechanistic model data and generates simulated data samples, while the discriminator judges the authenticity of the input data samples. This process of mutual adversarial interaction yields a large amount of near-realistic "synthetic" data.

[0099] (3) Pre-training

[0100] The neural network training process in this paper is based on the Adam algorithm for parameter optimization. Through iterative steps, the minimum loss function and optimal model weights are obtained. The algorithm initializes the parameters and weights at the start of each iteration. Therefore, we pre-train the CNN-GRU model using a large amount of near-realistic "synthetic" data generated by the aforementioned GANs data augmentation. Because the vortex expander model used comprehensively considers factors such as leakage and friction, pre-training can provide a more accurate and realistic initialization state.

[0101] 3. Embedding physics knowledge

[0102] To make the neural network modeling results more consistent with physical laws (energy conservation, mass conservation), it is necessary to construct physical constraints based on conservation conditions. This paper uses a loss function for knowledge embedding, that is, by adding physical equation constraints, the mechanism model guides the data model to model, with data-driven and knowledge-driven model enhancement, while avoiding inconsistencies in physical laws during modeling.

[0103] The loss function for embedding physical constraints is:

[0104] Loss = Loss GRU +λLoss PHYSICS (16)

[0105] In the formula, Loss represents the total loss of the model. GRU Loss is the standard loss of the data model (i.e., the supervised loss that measures the difference between the measured and predicted values). PHYSICS The loss is based on physical constraints and is weighted by the hyperparameter λ.

[0106] Based on the physical knowledge of vortex expanders, physical constraint loss functions for energy conservation and mass conservation laws are established respectively.

[0107] (1) Based on the law of conservation of mass, the mass flow constraint loss function of the vortex expander is established. During the operation of the vortex expander, the axial gap between the contact end faces of the moving and stationary vortex disks leads to radial leakage and axial leakage. At the same time, due to the back pressure, some gas will leak into the atmosphere.

[0108] Under ideal conditions, the average mass flow rate of a scroll compressor is:

[0109]

[0110] Radial leakage and axial leakage are respectively

[0111]

[0112]

[0113] In the formula σ x It is the axial leakage coefficient, σ y It is the radial leakage coefficient, γ x It is the axial leakage area, γ y It is the radial leakage area, h i and h o These are the gas enthalpy of the high-pressure side chamber and the low-pressure side chamber, respectively.

[0114] The leakage rate of the vortex expander per revolution is:

[0115]

[0116] Here, L is approximately a constant, and the leakage mass flow rate between the chambers is:

[0117]

[0118] According to the formula for calculating the mass flow rate of gas pressure difference flow, the leakage flow rate between the control volume chamber and the outside due to back pressure is:

[0119]

[0120] In the formula, γ is the leakage area of ​​the orifice, μ air is the coefficient of friction of the gas.

[0121] Therefore, the mass change and the loss function obtained based on the law of conservation of mass are as follows:

[0122]

[0123]

[0124] In the formula, the parameter q lim This is the threshold for the mass conservation loss, controlling the balance between the standard GRU loss and the mass conservation loss. This threshold is introduced because the physical process may be affected by less important unknown variables not included in the model, or by errors in experimental measurement data. A linear rectified activation function, ReLU(*), is used so that only differences exceeding the threshold are penalized. The model is updated using backpropagation with the Adam optimizer.

[0125] (2) Based on the law of conservation of energy, establish the output torque constraint loss function of the vortex expander. During the operation of the vortex expander, additional torque losses will be generated, including mechanical friction losses and viscous friction losses.

[0126] The output torque under ideal conditions is:

[0127]

[0128] The expressions for the mechanical friction loss term and the viscous friction torque are as follows:

[0129]

[0130]

[0131] Where, μ f It is the coefficient of friction, μ vf It is the coefficient of viscous friction, A fIt is the frictional contact area, A vf is the viscous friction contact area, and r is the base circle radius.

[0132] Therefore, the torque change and the loss function based on the law of energy conservation are obtained as follows:

[0133] ΔM=M+M f +M vf (28)

[0134]

[0135] In the formula, parameter M lim It is the threshold for energy conservation loss.

[0136] Combining equations (24) and (29), the loss function for physical constraints constructed in this paper can be expressed as:

[0137] Loss = Loss GRU +λ q Loss q +λ M Loss M (30)

[0138] In the formula: Loss GRU For purely data-driven neural network supervision loss; λ q and λ M These are the weighting coefficients for mass flow rate and output torque, respectively.

[0139] After pre-training the model using simulated data, physical constraints are added to the model training. Finally, the model is modeled using real experimental data. Since the model is based on mechanistic models and physical laws, physical consistency is guaranteed.

[0140] Thus, a physics-guided CNN-GRU hybrid neural network model was established. Using this hybrid model, the efficiency of the vortex expander was calculated based on the concept of air effective energy, and its efficiency was analyzed. The variation law of the vortex expander's efficiency was analyzed and summarized, providing guidance for its optimal control.

[0141] First, to intuitively evaluate the energy utilization and conversion capabilities of a scroll expander, it is necessary to select a suitable expansion efficiency evaluation index. This invention defines the expansion efficiency of a scroll expander as the ratio of the shaft output power Po to the aerodynamic power P of the compressed air. air The ratio, the expansion efficiency is expressed as:

[0142]

[0143] Using the established hybrid model of volumetric flow rate and output torque of the vortex expander, the expansion efficiency in the range of inlet pressure from 2.5 bar to 7 bar was simulated, and the variation law of expansion efficiency with inlet pressure, speed and inlet temperature was analyzed.

[0144] Analysis revealed the following trends in expansion efficiency with increasing inlet pressure and rotational speed: As rotational speed increases, the impact of leakage on the compression process gradually decreases, while the impact of friction on system efficiency becomes more apparent. Therefore, at constant pressure, the expansion efficiency is lower in the high and low speed operating ranges, with an efficiency extreme value appearing in the intermediate speed range. Under constant speed conditions, as the inlet pressure increases from 2.5 bar to 7 bar, the expansion efficiency exhibits a clear pattern of first increasing and then decreasing. These trends indicate that the scroll expander has a high-efficiency operating range.

[0145] Analysis of the expansion efficiency of the scroll expander shows that, for different load requirements, controlling the inlet pressure and speed of the scroll expander can make it operate in the high-efficiency range, thereby improving the energy utilization and conversion capability of the scroll expander.

[0146] Figure 4 The graph shows the expansion efficiency as a function of rotational speed under different inlet pressures. It can be seen from the graph that when the inlet pressure of the vortex expander is constant, there is an optimal rotational speed that maximizes the expansion efficiency.

[0147] from Figure 4 It can also be seen that as the intake pressure increases, the expansion efficiency also increases, and the optimal speed shifts towards higher speeds. The curve connecting the points of maximum efficiency under different intake pressures represents the optimal operating trajectory of the scroll expander. Based on different system efficiency optimization control requirements, this optimal operating curve can determine the current optimal combination of intake pressure and speed, thereby controlling the system to operate in the high-efficiency zone.

[0148] Analysis using a hybrid model of volumetric flow rate and output torque shows that increased inlet temperature results in a smaller volumetric flow rate of the high-pressure gas converted to standard conditions, while higher inlet temperature leads to greater output power from the vortex expander, thus improving expansion efficiency. Therefore, the rational recovery and utilization of heat and cold generated by compressed air energy storage systems can effectively improve system efficiency and has been applied in advanced adiabatic compressed air energy storage systems.

[0149] Data acquisition and experimental verification for modeling a vortex expander both require the support of an actual system. Therefore, in order to verify the effectiveness of the model established in this invention, a vortex expander experimental system was built. Experimental data was collected through multiple sensors to verify the rationality and accuracy of the established hybrid model.

[0150] (1) Vortex expander experimental system

[0151] The structure of the experimental system is shown in Figure 5(a). The main hardware equipment includes key equipment such as gas tank, pressure regulating valve, gas heater, vortex expander, permanent magnet synchronous generator, and power electronic frequency converter, as well as pressure, flow, temperature and speed / torque sensors. The physical distribution of the experimental platform is shown in Figure 5(b).

[0152] The operating principle of the experimental system is as follows:

[0153] The pressure regulating valve at the gas tank outlet controls the inlet pressure of the scroll expander, while the solenoid valve is a switching device that controls the on / off state of the gas path. A heater is connected to the inlet of the scroll expander, and its temperature control system controls the inlet temperature, simulating the preheating process of the inlet gas in a heat storage device. High-pressure gas enters the scroll expander, driving a permanent magnet synchronous generator coaxially connected to it to complete the power generation process. An uncontrolled rectifier circuit and a BUCK circuit are connected after the generator, with a 180V battery as the load. The generator speed can be adjusted by changing the duty cycle of the switching devices in the BUCK circuit.

[0154] The experimental data collection and processing are as follows:

[0155] The experimental system's gas path section is equipped with pressure sensors to collect throttling pressure and scroll expander inlet pressure data. In addition, the gas path section also includes a flow meter to collect the scroll expander inlet volumetric flow rate, and a temperature sensor to measure the inlet temperature. The system's mechanical connections are fitted with speed / torque sensors to collect scroll expander speed and output torque data.

[0156] Experimental data were collected using a built micro compressed air energy storage and power generation system. Within the range of 2.5 bar to 4.0 bar inlet pressure of the vortex expander, experimental data such as rotational speed, inlet temperature, volumetric flow rate, and output torque were collected.

[0157] (2) Training results of the hybrid model

[0158] The rationality and accuracy of the proposed mechanism- and data-driven integrated energy system underlying equipment modeling method were evaluated using the aforementioned vortex expander experimental system.

[0159] The modeling results are as follows:

[0160] A neural network model of the vortex expander was obtained by training it using experimental data. The root mean square error variation during the training process is shown below. Figure 6 As shown.

[0161] The verification results of the hybrid model of the vortex expander are shown in Figures 7(a) and 7(b). The red circle in the figure represents the output value of the mathematical model, and the blue line represents the actual experimental data.

[0162] The established volumetric flow rate model has a mean absolute error (MAPE) of 0.91% and a root mean square error (RMSE) of 3.4511, indicating that the error of the volumetric flow rate hybrid model is extremely small. Similarly, the established output torque model has a MAPE of 3.30% and an RMSE of 0.2301. These results show that the relative errors between the output torque and volumetric flow rate of the vortex expander hybrid model and the actual values ​​are very small, ensuring the modeling accuracy.

[0163] To verify the accuracy and effectiveness of the model established in this invention, the following two schemes were designed for horizontal and vertical verification comparison:

[0164] Longitudinal comparison: The modeling accuracy of the non-physically guided CNN-GRU hybrid neural network method is compared with that of the proposed method. The comparison results are shown in Table 1:

[0165] Table 1 Comparison of modeling errors of different methods

[0166]

[0167] As shown in Table 1, by comparing the mean absolute error percentage (MAPE) and root mean square error (RMSE) of modeling based on physical guidance, it is clear that after adding the pre-training part of the mechanistic model, the MAPE decreased from 1.41% to 0.88%, and the RMSE decreased from 4.8465 to 3.2531.

[0168] Therefore, compared with the model built by the simple hybrid neural network CNN-GRU modeling method, the model built using the physics-guided data modeling method proposed in this paper has smaller errors and higher modeling accuracy.

[0169] Horizontal Comparison: The model established by the hybrid modeling method of this invention is compared horizontally with neural network modeling methods such as RNN, CNN, LSTM, GRU, and CNN-LSTM. The prediction errors and modeling times of different methods are compared to examine their accuracy. The results are as follows: Figure 8 As shown.

[0170] pass Figure 8 The radar chart comparison shows that CNN-LSTM and CNN-GRU methods have higher modeling accuracy and smaller errors compared to RNN, CNN, LSTM, GRU and other methods.

[0171] Furthermore, the running times for CNN-LSTM and CNN-GRU methods are 22 minutes 43 seconds and 11 minutes 24 seconds, respectively. It can be seen that the CNN-GRU method is more efficient while improving modeling accuracy.

[0172] The results of the above verification experiments show that the physics-guided CNN-GRU hybrid neural network modeling method proposed in this embodiment has smaller model errors and higher modeling accuracy compared with traditional neural network modeling methods, indicating that the established model can more accurately predict device performance.

[0173] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A physical-guided and data-driven integrated energy system equipment modeling method, characterized in that, Includes the following steps: By combining convolutional neural networks and gated recurrent unit neural networks, a hybrid data-driven model is formed. When training the hybrid data-driven model, the Adam algorithm is used to iteratively update the weights. The weights and biases of each neuron are continuously updated through momentum and adaptive learning rate, so that the output value of the loss function reaches the optimal value. A mechanism model of integrated energy system equipment is constructed, data augmentation of the mechanism model is performed, simulation data is generated, and the simulation data is used to pre-train the hybrid data-driven model. Generative adversarial networks (GANs) are used to augment the data of a mechanistic model. The GAN includes a generator and a discriminator. The generator receives random variables and generates data samples. The discriminator determines whether the input samples are real or synthetic. The discriminator also calculates the loss and performs backpropagation to find the gradient, thereby updating the parameters. Based on the physical laws of the integrated energy system equipment, physical constraints are constructed and modeled into the loss function of the hybrid data-driven model. The hybrid data-driven model is trained using actual data to obtain the final modeling result. Based on the physical laws of the integrated energy system equipment, the specific process of constructing physical constraints includes constructing physical constraints with energy conservation and / or mass conservation, embedding constraints through a loss function, and adjusting the proportion of constraints in the overall loss function using weighting coefficients. (1) Based on the law of conservation of mass, the mass flow constraint loss function of the vortex expander is established. During the operation of the vortex expander, the axial gap between the contact end faces of the moving and stationary vortex disks leads to radial leakage and axial leakage. At the same time, due to the back pressure, some gas will leak into the atmosphere. Under ideal conditions, the average mass flow rate of a scroll compressor is: （17） Radial leakage and axial leakage are respectively （18） （19） In the formula It is the axial leakage coefficient. It is the radial leakage coefficient. It is the axial leakage area. It is the radial leakage area. and These are the gas enthalpies of the high-pressure side chamber and the low-pressure side chamber, respectively. The leakage rate of the vortex expander per revolution is: （20） The leakage mass flow rate between chambers is: （21） According to the formula for calculating the mass flow rate of gas pressure difference flow, the leakage flow rate between the control volume chamber and the outside due to back pressure is: （22） In the formula, Let be the leakage area of ​​the orifice. The coefficient of friction of the gas; Therefore, the mass change and the loss function obtained based on the law of conservation of mass are as follows: （23） （24） In the formula, the parameter This is the threshold for the mass conservation loss, which controls the balance between the standard GRU loss and the mass conservation loss. A linear rectified activation function ReLU (*) is used so that only differences greater than the threshold are penalized. Backpropagation of the Adam optimizer is used to update the model. (2) Based on the law of conservation of energy, establish the output torque constraint loss function of the vortex expander; during the operation of the vortex expander, additional torque loss will be generated, including mechanical friction loss and viscous friction loss; The output torque under ideal conditions is: （25） The expressions for the mechanical friction loss term and the viscous friction torque are as follows: （26） （27） in, It is the coefficient of friction. It is the coefficient of viscous friction. It is the frictional contact area. It is the viscous friction contact area. It is the base circle radius; Therefore, the torque change and the loss function based on the law of energy conservation are obtained as follows: （28） （29） In the formula, the parameter It is the threshold for energy conservation loss; The loss function for the constructed physical constraints can be expressed as: （30） In the formula: For purely data-driven neural network supervision loss; and These are the weighting coefficients for mass flow rate and output torque, respectively. To establish the mass flow constraint loss function for the vortex expander based on the law of conservation of mass; Based on the law of conservation of energy, the output torque constraint loss function of the vortex expander is established as the energy conservation loss; When constructing the mechanism model of integrated energy system equipment, the effects of leakage, over- and under-expansion and friction factors are comprehensively considered to establish mechanism models for volumetric flow rate and output torque respectively. Rotational speed, intake pressure and / or temperature are used as control variables to establish mathematical models of volumetric flow rate and / or output torque with respect to the control variables.

2. The physical-guided and data-driven integrated energy system equipment modeling method as described in claim 1, characterized in that, The hybrid data-driven model includes an input layer, a CNN layer, a GRU layer, a fully connected layer, and an output layer set up in one step. The CNN layer includes several convolutional layers and pooling layers that are stacked alternately, and there is at least one GRU layer.

3. The physical-guided and data-driven integrated energy system equipment modeling method as described in claim 2, characterized in that, The convolutional layer is a one-dimensional convolution, with the convolution method being Same convolution and the activation function being ReLU; the pooling layer uses Valid max pooling, and the fully connected layer uses the Sigmoid activation function.

4. A physical-guided and data-driven integrated energy system equipment modeling system, based on the physical-guided and data-driven integrated energy system equipment modeling method as described in any one of claims 1-3, characterized in that, include: The hybrid data-driven model building module is configured to combine convolutional neural networks and gated recurrent unit neural networks to form a hybrid data-driven model. When training the hybrid data-driven model, the Adam algorithm is used to iteratively update the weights. The weights and biases of each neuron are continuously updated through momentum and adaptive learning rate, so that the output value of the loss function reaches the optimal value. The pre-training fusion module is configured to construct a mechanism model of integrated energy system equipment, perform data augmentation on the mechanism model, generate simulation data, and use the simulation data to pre-train the hybrid data-driven model. Generative adversarial networks (GANs) are used to augment the data of a mechanistic model. The GAN includes a generator and a discriminator. The generator receives random variables and generates data samples. The discriminator determines whether the input samples are real or synthetic. The discriminator also calculates the loss and performs backpropagation to find the gradient, thereby updating the parameters. The training fusion module is configured to construct physical constraints based on the physical laws of the integrated energy system equipment, and model them into the loss function of the hybrid data-driven model. The hybrid data-driven model is then trained using real data to obtain the final modeling result.

5. A computer-readable storage medium, characterized in that, It stores multiple instructions, which are adapted to be loaded by the processor of the terminal device and executed as steps in the method of any one of claims 1-3.

6. A terminal device, characterized in that, Including processors and computer-readable storage media The processor is used to implement the instructions; the computer-readable storage medium is used to store a plurality of instructions adapted to be loaded by the processor and executed by the steps of the method according to any one of claims 1-3.

Images