Current sharing control method and system for an excitation system
By using a temperature prediction model based on current-temperature dual-parameter feedback and an LSTM network, the overheating problem caused by uneven temperature in the excitation system is solved, achieving high-precision current sharing control, extending device life, optimizing energy consumption, and improving system stability.
Patent Information
- Application Number
- CN202610412799.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-31
- Publication Date
- 2026-06-26
AI Technical Summary
Existing current sharing control methods for excitation systems rely on feedback from a single current parameter, ignoring temperature parameters. This leads to uneven temperature distribution, causing localized overheating, thermal fatigue, and device failure. Furthermore, the dynamic adaptability is insufficient, failing to achieve thermal-electric synergistic optimization and increasing energy consumption.
A temperature prediction model is constructed by adopting a current-temperature dual-parameter feedback mechanism and combining it with an LSTM recurrent neural network. By collecting current and temperature data in real time, trigger pulse adjustment commands are generated to achieve high-precision current sharing control between rectifier cabinets and between thyristors.
It achieves high-precision current sharing between rectifier cabinets and between thyristors, avoiding local overheating, extending device life, reducing energy consumption, and improving system stability and dynamic adaptability.
Smart Images
Figure CN122292857A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of excitation system control technology, and specifically relates to a current sharing control method and system for an excitation system. Background Technology
[0002] In excitation systems, current sharing control mainly includes cabinet current sharing (current sharing between rectifier cabinets) and tube current sharing (current sharing between thyristors). Cabinet current sharing is typically achieved by collecting the output current of parallel rectifier cabinets, comparing current deviations, and then adjusting the trigger pulse angle (e.g., delaying or advancing) or load distribution coefficient to balance the output current of each cabinet. For example, in a digital current sharing module, after calculating the current difference between cabinets, the trigger pulse angle is dynamically modified or a current sharing reactor is used to achieve current balancing.
[0003] Current sharing schemes focus on multiple thyristors on the same bridge arm. By collecting the real-time current values of each thyristor and comparing their deviations, the trigger pulse angle or gate drive signal of each thyristor is fine-tuned to balance the current distribution. These methods generally rely on single-parameter current feedback and achieve current sharing through closed-loop control algorithms (such as PID or comparator logic).
[0004] The existing technology has the following drawbacks: Single-parameter dependency risk: Existing current sharing control relies solely on current parameters as feedback, ignoring the influence of temperature parameters. Temperature changes in the thyristor (such as increased case temperature) directly affect its conduction characteristics, thermal stress, and aging rate. Under current sharing conditions, uneven temperature distribution may lead to localized overheating, thermal fatigue, or even device failure, shortening equipment lifespan.
[0005] Insufficient dynamic adaptability: Although current parameters can reflect instantaneous load distribution, they cannot capture the cumulative effect of temperature. When there are sudden load changes or changes in ambient temperature, pure current sharing may lead to a deterioration of temperature distribution, causing fluctuations in the current sharing coefficient and reducing system stability.
[0006] Energy efficiency optimization deficiencies: Ignoring temperature parameters prevents the achievement of synergistic thermo-electric optimization, for example... Figure 1 The junction temperature and current curves of the device shown indicate that the on-state current of the thyristor decreases as the temperature increases. However, pure current sharing may overload the high-temperature device, increasing energy consumption and heat dissipation burden. Summary of the Invention
[0007] To address the aforementioned issues, this application provides a current sharing control method and system for an excitation system. By utilizing a current-temperature dual-parameter feedback mechanism and the dynamic temperature prediction capability of an LSTM recurrent neural network, it can overcome the limitations of a single parameter in traditional pure current sharing, taking into account both instantaneous current balance and long-term thermal stress optimization, thereby achieving high-precision current sharing between rectifier cabinets and between thyristors.
[0008] The following is the technical content of this disclosure: A current sharing control method for an excitation system, characterized in that it includes: The system acquires the real-time current values of each thyristor in the excitation system and the output current value of the rectifier cabinet, forms a historical current sequence, and calculates the current change rate; at the same time, it collects the thyristor case temperature sequence, the heat sink temperature sequence, and the ambient temperature sequence in real time. A temperature prediction model is constructed, which is used to output the future temperature sequence of the thyristor based on a multi-dimensional feature vector constructed from the thyristor case temperature sequence, heat sink temperature sequence, ambient temperature sequence, historical current sequence and current change rate. Based on the future temperature sequence of the thyristor output by the temperature prediction model, and combined with the real-time current value of the thyristor and the output current value of the rectifier cabinet, a trigger pulse adjustment command is generated to realize the control of current sharing between rectifier cabinets and between thyristors.
[0009] Furthermore, The real-time current is acquired by a Hall sensor, and the thyristor case temperature, radiator temperature, and ambient temperature are acquired by a thermocouple or an infrared sensor.
[0010] Furthermore, Before constructing the temperature prediction model, the method further includes: constructing an electrothermal coupled state-space model to quantify the dynamic coupling relationship between current parameters and temperature parameters, providing constraints for the prediction results of the temperature prediction model; The electrothermal coupling state-space model is as follows:
[0011] in, To control the cycle; , u is the ideal no-load DC voltage of the rectifier. k Control quantity related to firing angle The current value of the kth excitation branch at the kth sampling time; L is the equivalent inductance value of the corresponding branch in the excitation system; The grid-side voltage at the excitation system connection point at the k-th sampling time; For the k-th sampling time, the voltage drop across the resistive element in the excitation circuit; For the k-th sampling time, the real-time temperature value of the corresponding branch in the excitation system; It is the thermal resistance coefficient; The heat dissipation coefficient; The ambient temperature value of the environment in which the excitation system is located.
[0012] Furthermore, The temperature prediction model is an LSTM recurrent neural network, and the training process of the LSTM recurrent neural network includes: Construct a multidimensional feature vector, wherein the multidimensional feature vector is:
[0013] in, (t) represents the thyristor case temperature at time t; (t) represents the radiator temperature at time t; (t) represents the ambient temperature at time t; (t) represents the current at time t; Let be the rate of change of current at time t, where Let be the current value of the k-th excitation branch at time t. For the k-th excitation branch in The current value at time [time]. To control the cycle; A training set is constructed using the multidimensional feature vectors. The mean vector and standard deviation vector of the training set are calculated. The multidimensional feature vectors are then standardized based on the mean vector and standard deviation vector to obtain standardized feature vectors. Set a window length L1, generate input-output pairs, where the input is a standardized feature vector segment of length L1, and the output is the temperature sequence of the corresponding time period; The input-output pair is fed into an LSTM network with a single hidden layer structure for training, and the training objective of the LSTM network is to minimize the mean squared error.
[0014] Furthermore, The prediction process of the temperature prediction model includes: Starting from the window of data at the end of the training set, the future temperature sequence is predicted iteratively through the trained LSTM network; The predicted results are de-standardized to restore the temperature dimension, and the restored predicted temperature is used to replace the measured temperature in control decisions.
[0015] Furthermore, Generating the trigger pulse adjustment command includes: constructing an objective function, obtaining the trigger angle correction amount by minimizing the objective function, and generating the trigger pulse adjustment command based on the trigger angle correction amount; The objective function is the weighted combined value of current equalization deviation and temperature equalization deviation; The objective function is:
[0016] in, Let be the current value of the k-th excitation branch at time (k+1), calculated by discretization of the electrical equations; This represents the average value of the bridge arm current; The temperature value at time (k+1) is output by the LSTM network; For preset temperature, Fixed current weighting; In the formula:
[0017] The temperature-adaptive weighted baseline value. Where K is the temperature threshold and K is the gain coefficient. The predicted temperature of the k-th thyristor; The minimization objective function is:
[0018] in, For the k-th switching function, To make J reach its minimum value The value of ; The trigger angle correction amount is:
[0019] in It is an adaptive learning rate.
[0020] Furthermore, The gradient descent method is used to solve the minimization objective function. .
[0021] Furthermore, Calculated using the finite difference method .
[0022] A current sharing control system for an excitation system, comprising: The data acquisition module is used to acquire the real-time current values of each thyristor in the excitation system and the output current value of the rectifier cabinet, form a historical current sequence and calculate the current change rate; at the same time, it acquires the thyristor case temperature sequence, heat sink temperature sequence and ambient temperature sequence in real time. The model building module is used to build a temperature prediction model. The temperature prediction model is used to build a multi-dimensional feature vector based on the thyristor case temperature sequence, heat sink temperature sequence, ambient temperature sequence, historical current sequence and current change rate, and output the future temperature sequence of the thyristor. The control module is used to generate trigger pulse adjustment commands based on the future temperature sequence of the thyristors output by the temperature prediction model, combined with the real-time current value of the thyristors and the output current value of the rectifier cabinet, so as to realize the control of current sharing between rectifier cabinets and between thyristors.
[0023] Furthermore, The temperature prediction model is an LSTM recurrent neural network, and the training process of the LSTM recurrent neural network includes: Construct a multidimensional feature vector, wherein the multidimensional feature vector is:
[0024] in, (t) represents the thyristor case temperature at time t; (t) represents the radiator temperature at time t; (t) represents the ambient temperature at time t; (t) represents the current at time t; Let be the rate of change of current at time t, where Let be the current value of the k-th excitation branch at time t. For the k-th excitation branch in The current value at time [time]. To control the cycle; A training set is constructed using the multidimensional feature vectors. The mean vector and standard deviation vector of the training set are calculated. The multidimensional feature vectors are then standardized based on the mean vector and standard deviation vector to obtain standardized feature vectors. Set a window length L1, generate input-output pairs, where the input is a standardized feature vector segment of length L1, and the output is the temperature sequence of the corresponding time period; The input-output pair is fed into an LSTM network with a single hidden layer structure for training, and the training objective of the LSTM network is to minimize the mean squared error.
[0025] Compared with the prior art, this application has the following advantages: This application collects the current timing characteristics (real-time current, historical current sequence, current change rate) and multi-dimensional temperature timing characteristics (thyristor case temperature, heat sink temperature, ambient temperature) of the excitation system thyristor and rectifier cabinet. It uses LSTM neural network to mine the timing correlation law of electrothermal parameters to build a temperature prediction model to predict the future temperature rise trend of the thyristor. Then, it combines the real-time current monitoring value and temperature prediction results to carry out current-temperature dual-target regulation and generate trigger pulse adjustment commands to realize active predictive current sharing control between rectifier cabinets and thyristors. By integrating the instantaneous dynamic changes of electrical parameters with the cumulative changes of thermal parameters, the temperature rise pattern of the thyristor can be predicted in advance through time-series forecasting. This allows current sharing control to not only focus on the instantaneous balance of current, but also to intervene in load distribution in advance based on temperature development trends. At the same time, it achieves synergistic optimization of electrical and thermal parameters to rationally distribute the load and avoid excessive load on high-temperature devices. Therefore, it can solve the problems of local overheating and device failure caused by pure current single-parameter dependence, and the problems of current sharing coefficient fluctuation and insufficient system dynamic adaptability when the temperature cumulative effect causes sudden load changes or changes in ambient temperature.
[0026] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0028] Figure 1 A schematic diagram of the method of this disclosure is shown; Figure 2 A schematic diagram of junction temperature and current curves according to an embodiment of this disclosure is shown; Figure 3 A schematic diagram of an LSTM-based temperature prediction framework according to an embodiment of this disclosure is shown. Detailed Implementation
[0029] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0030] like Figure 1 This is a schematic diagram of the method disclosed herein. The disclosure proposes a closed-loop current sharing control method that integrates current and temperature dual-parameter feedback. The core of this method lies in real-time acquisition of current and temperature data, and dynamic adjustment of control commands through a dual-parameter optimization algorithm to achieve coordinated optimization of cabinet current sharing and pipe current sharing. Specific solutions include: 1. Step One: Data Acquisition Design Each rectifier cabinet is equipped with a current sensor (such as a Hall effect sensor) and a temperature sensor (such as a thermocouple or infrared sensor). The current sensor collects the real-time current value of each thyristor (for tube current sharing) and the output current of the rectifier cabinet (for cabinet current sharing); the temperature sensor monitors the thyristor case temperature, heat sink temperature, and ambient temperature. The data is transmitted to the central processing unit or distributed control unit via a high-speed communication network.
[0031] 2. Step Two: Constructing a Temperature Prediction Model 2.1) Establish the discretized electro-thermal dynamic equations for the thyristor bridge arms and construct an electro-thermal coupled state-space model (e.g., Figure 2 (Schematic diagram of junction temperature versus current curves) The electrothermal coupling state-space model is as follows:
[0032] in, To control the cycle; , u is the ideal no-load DC voltage of the rectifier. k Control quantity related to firing angle The current value of the kth excitation branch at the kth sampling time; L is the equivalent inductance value of the corresponding branch in the excitation system; The grid-side voltage at the excitation system connection point at the k-th sampling time; For the k-th sampling time, the voltage drop across the resistive element in the excitation circuit; For the k-th sampling time, the real-time temperature value of the corresponding branch in the excitation system; It is the thermal resistance coefficient; The heat dissipation coefficient; This refers to the ambient temperature of the environment in which the excitation system is located.
[0033] 2.2) Dynamic prediction of temperature changes is achieved using an LSTM recurrent neural network, specifically as follows: 1. Training Phase: Acquiring the SCR case temperature sequence Radiator temperature sequence Ambient temperature sequence Historical current sequence And the rate of change of current, construct a multidimensional feature vector: (2) in, (t) represents the thyristor case temperature at time t; (t) represents the radiator temperature at time t; (t) represents the ambient temperature at time t; (t) represents the current at time t; Let be the rate of change of current at time t, where is the current value of the k-th excitation branch at time t, is the current value of the k-th excitation branch at time; is the control period; Calculate the mean vector (m < t) of the training set and the standard deviation vector , and standardize the feature vector: (3) Set the window length L1 and generate input-output pairs Train the LSTM network, that is, the input of the segmented model is and the corresponding theoretical output is .
[0034] Furthermore, input into the hidden layer. The LSTM network adopts a single hidden layer structure, and the hidden layer is composed of L1 isomorphic LSTM cells connected in series, and the size of the state vector of each cell is , and its forward calculation follows the gating mechanism: (4) where are the input gate, forget gate, cell state, and output gate respectively; is the forward activation function; represents element-wise multiplication. x t is the input vector at the t-th time step; h t 1 is the hidden state vector at the t 1-th time step; is the weight coefficient matrix (W xi , W hi , W ci are the weight matrices from the input x t , the previous hidden state h t 1, the previous cell state c t 1 to the input gate i t ; W xf , W hf , W cf are the weight matrices from the input x t , the previous hidden state h t 1, the previous cell state c t 1 to the forget gate f t ; W xc , W hc: These are the inputs x t Previous hidden state h t The weight matrix from 1 to candidate cell states; W xo W ho W co : These are the inputs x t Previous hidden state h t 1. Previous cell state c t 1 to output gate o t (weight matrix) For the bias vector (b) i b f b c b o These are the bias vectors for the input gate, forget gate, candidate cell states, and output gate, respectively.
[0035] but The output after the hidden layer can be represented as: (5) In the formula, and These represent the state and output of the previous LSTM cell, respectively; LSTMforward represents the forward computation method of the LSTM cell (Equation (3)).
[0036] The training process aims to minimize the mean squared error (MSE). (6) Once the training objective is achieved, the final hidden network can be obtained.
[0037] 2. Prediction Phase: An iterative method is used to generate future temperature sequences. Starting from the end of the training set window... Starting with the trained model Iterative prediction: (7) Finally, the temperature dimension is reduced by denormalization: And will predict temperature Instead of using measured temperature in control decisions, Synchronized with the control system cycle.
[0038] like Figure 3 This is a temperature prediction framework based on LSTM.
[0039] 3. Step Three: Generate control commands based on the predicted temperature and measured current. The specific algorithm is as follows: Construct the minimum objective function Dynamically generate start pulse adjustment commands: (8) in, Let be the current value of the k-th excitation branch at time (k+1), calculated by discretization of the electrical equations; This represents the average value of the bridge arm current; The temperature value at time (k+1) is output by the LSTM network; For preset temperature, To fix the current weight and ensure the current sharing performance of the foundation, the formula is: (9) The temperature-adaptive weighted baseline value. Where K is the temperature threshold and K is the gain coefficient. Let be the predicted temperature of the k-th thyristor.
[0040] The optimal switching function is solved using the gradient descent method. : (10) in, For the k-th switching function, To make J reach its minimum value The value of ; Finally, the trigger angle correction amount is obtained: (11) in The adaptive learning rate is calculated using the finite difference method. .
[0041] In summary: This technical solution eliminates the blind zone of pure current sharing through dual-parameter feedback of current and temperature, further optimizing the current sharing coefficient and reducing thermal stress failures. Its temperature equalization control prevents localized overheating and extends the lifespan of the silicon controlled rectifier (SCR); collaborative optimization reduces heat dissipation energy consumption. Simultaneously, by updating input data in real time through a sliding window, it can adapt to nonlinear temperature rises caused by sudden load changes.
[0042] Based on the above method, this disclosure also provides a current sharing control system for an excitation system, comprising: The data acquisition module is used to acquire the real-time current values of each thyristor in the excitation system and the output current value of the rectifier cabinet, form a historical current sequence and calculate the current change rate; at the same time, it acquires the thyristor case temperature sequence, heat sink temperature sequence and ambient temperature sequence in real time. The model building module is used to build a temperature prediction model. The temperature prediction model is used to build a multi-dimensional feature vector based on the thyristor case temperature sequence, heat sink temperature sequence, ambient temperature sequence, historical current sequence and current change rate, and output the future temperature sequence of the thyristor. The control module is used to generate trigger pulse adjustment commands based on the future temperature sequence of the thyristors output by the temperature prediction model, combined with the real-time current value of the thyristors and the output current value of the rectifier cabinet, so as to realize the control of current sharing between rectifier cabinets and between thyristors.
[0043] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A current sharing control method for an excitation system, characterized in that, include: The system acquires the real-time current values of each thyristor in the excitation system and the output current value of the rectifier cabinet, forms a historical current sequence, and calculates the current change rate; at the same time, it collects the thyristor case temperature sequence, the heat sink temperature sequence, and the ambient temperature sequence in real time. A temperature prediction model is constructed, which is used to output the future temperature sequence of the thyristor based on a multi-dimensional feature vector constructed from the thyristor case temperature sequence, heat sink temperature sequence, ambient temperature sequence, historical current sequence and current change rate. Based on the future temperature sequence of the thyristor output by the temperature prediction model, and combined with the real-time current value of the thyristor and the output current value of the rectifier cabinet, a trigger pulse adjustment command is generated to realize the control of current sharing between rectifier cabinets and between thyristors.
2. The current sharing control method for the excitation system according to claim 1, characterized in that, The real-time current is acquired by a Hall sensor, and the thyristor case temperature, radiator temperature, and ambient temperature are acquired by a thermocouple or an infrared sensor.
3. The current sharing control method for the excitation system according to claim 1, characterized in that, Before constructing the temperature prediction model, the method further includes: constructing an electrothermal coupled state-space model to quantify the dynamic coupling relationship between current parameters and temperature parameters, providing constraints for the prediction results of the temperature prediction model; The electrothermal coupling state-space model is as follows: in, To control the cycle; , u is the ideal no-load DC voltage of the rectifier. k Control quantity related to firing angle The current value of the kth excitation branch at the kth sampling time; L is the equivalent inductance value of the corresponding branch in the excitation system; The grid-side voltage at the excitation system connection point at the k-th sampling time; For the k-th sampling time, the voltage drop across the resistive element in the excitation circuit; For the k-th sampling time, the real-time temperature value of the corresponding branch in the excitation system; It is the thermal resistance coefficient; The heat dissipation coefficient; This refers to the ambient temperature of the environment in which the excitation system is located.
4. The current sharing control method for the excitation system according to claim 1, characterized in that, The temperature prediction model is an LSTM recurrent neural network, and the training process of the LSTM recurrent neural network includes: Construct a multidimensional feature vector, wherein the multidimensional feature vector is: in, (t) represents the thyristor case temperature at time t; (t) represents the radiator temperature at time t; (t) represents the ambient temperature at time t; (t) represents the current at time t; Let be the rate of change of current at time t, where Let be the current value of the k-th excitation branch at time t. For the k-th excitation branch in The current value at time [time]. To control the cycle; A training set is constructed using the multidimensional feature vectors. The mean vector and standard deviation vector of the training set are calculated. The multidimensional feature vectors are then standardized based on the mean vector and standard deviation vector to obtain standardized feature vectors. Set a window length L1, generate input-output pairs, where the input is a standardized feature vector segment of length L1, and the output is the temperature sequence of the corresponding time period; The input-output pair is fed into an LSTM network with a single hidden layer structure for training, and the training objective of the LSTM network is to minimize the mean squared error.
5. The current sharing control method for the excitation system according to claim 4, characterized in that, The prediction process of the temperature prediction model includes: Starting from the window of data at the end of the training set, the future temperature sequence is predicted iteratively through the trained LSTM network; The predicted results are de-standardized to restore the temperature dimension, and the restored predicted temperature is used to replace the measured temperature in control decisions.
6. The current sharing control method for the excitation system according to claim 1, characterized in that, Generating the trigger pulse adjustment command includes: constructing an objective function, obtaining the trigger angle correction amount by minimizing the objective function, and generating the trigger pulse adjustment command based on the trigger angle correction amount; The objective function is the weighted combined value of current equalization deviation and temperature equalization deviation; The objective function is: in, Let be the current value of the k-th excitation branch at time (k+1), calculated by discretization of the electrical equations; This represents the average value of the bridge arm current; The temperature value at time (k+1) is output by the LSTM network; For preset temperature, Fixed current weighting; In the formula: The temperature-adaptive weighted baseline value. Where K is the temperature threshold and K is the gain coefficient. The predicted temperature of the k-th thyristor; The minimization objective function is: in, For the k-th switching function, To make J reach its minimum value The value of ; The trigger angle correction amount is: in It is an adaptive learning rate.
7. The current sharing control method for the excitation system according to claim 6, characterized in that, The gradient descent method is used to solve the minimization objective function. .
8. The current sharing control method for the excitation system according to claim 7, characterized in that, Calculated using the finite difference method .
9. A current sharing control system for an excitation system, characterized in that, include: The data acquisition module is used to acquire the real-time current values of each thyristor in the excitation system and the output current value of the rectifier cabinet, form a historical current sequence and calculate the current change rate; at the same time, it acquires the thyristor case temperature sequence, heat sink temperature sequence and ambient temperature sequence in real time. The model building module is used to build a temperature prediction model. The temperature prediction model is used to build a multi-dimensional feature vector based on the thyristor case temperature sequence, heat sink temperature sequence, ambient temperature sequence, historical current sequence and current change rate, and output the future temperature sequence of the thyristor. The control module is used to generate trigger pulse adjustment commands based on the future temperature sequence of the thyristors output by the temperature prediction model, combined with the real-time current value of the thyristors and the output current value of the rectifier cabinet, so as to realize the control of current sharing between rectifier cabinets and between thyristors.
10. The system according to claim 9, characterized in that, The temperature prediction model is an LSTM recurrent neural network, and the training process of the LSTM recurrent neural network includes: Construct a multidimensional feature vector, wherein the multidimensional feature vector is: in, (t) represents the thyristor case temperature at time t; (t) represents the radiator temperature at time t; (t) represents the ambient temperature at time t; (t) represents the current at time t; Let be the rate of change of current at time t, where Let be the current value of the k-th excitation branch at time t. For the k-th excitation branch in The current value at time [time]. To control the cycle; A training set is constructed using the multidimensional feature vectors. The mean vector and standard deviation vector of the training set are calculated. The multidimensional feature vectors are then standardized based on the mean vector and standard deviation vector to obtain standardized feature vectors. Set a window length L1, generate input-output pairs, where the input is a standardized feature vector segment of length L1, and the output is the temperature sequence of the corresponding time period; The input-output pair is fed into an LSTM network with a single hidden layer structure for training, and the training objective of the LSTM network is to minimize the mean squared error.