Intelligent control system for molten salt heating based on machine learning
By using a machine learning-based intelligent control system for molten salt heating, and employing an LSTM model and an adaptive fuzzy PID controller, accurate prediction and stable control of molten salt temperature are achieved. This solves the problems of slow response and high energy consumption in molten salt heating systems, and improves the stability and safety of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA RESOURCES POWER HEZE
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-02
AI Technical Summary
Existing molten salt heating control systems are slow to respond to large time delays and nonlinear conditions, resulting in large temperature fluctuations, high energy consumption, and frequent heater start-ups and shutdowns, which affect energy storage efficiency and equipment safety.
A machine learning-based intelligent control system for molten salt heating is adopted. The system acquires multi-dimensional data in real time through a data acquisition module, uses an LSTM model to predict the molten salt temperature, and combines it with an adaptive fuzzy PID controller to achieve feedforward control, smoothly adjust the heater power, and avoid temperature overshoot and oscillation.
It achieves precise tracking of molten salt temperature, reduces energy consumption, improves system robustness and safety, avoids drastic temperature fluctuations, and extends equipment life.
Smart Images
Figure CN122131666A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of thermal energy storage and intelligent control technology, and in particular to a machine learning-based intelligent control system for molten salt heating. Background Technology
[0002] Molten salt energy storage technology, as a highly efficient sensible heat energy storage method, has broad application prospects in fields such as solar thermal power generation, grid peak shaving and valley filling, and industrial off-peak electricity heating. In molten salt energy storage systems, the molten salt heating system is one of the core components. Molten salt is usually solid at room temperature and needs to be heated above its melting point and maintained within a certain high-temperature operating range (e.g., 250℃-550℃) by an electric heater or heat exchanger. Due to the large specific heat capacity of molten salt, its significant viscosity variation with temperature, and the fact that the heating system often has a large volume and complex piping structure, the entire heating process exhibits extremely strong nonlinearity, large time delay, and strong coupling thermodynamic characteristics.
[0003] In existing molten salt heating control applications, industry commonly employs traditional PID control algorithms or their simpler modifications to regulate heater power. The control system obtains the real-time temperature using temperature sensors placed within the molten salt tank or pipeline, compares it to the set target temperature, calculates the control input using a PID algorithm, and then adjusts the thyristor voltage regulator or solid-state relay to change the heater's supply voltage or duty cycle. This error feedback-based control method, when faced with changes in system operating conditions (such as sudden changes in flow rate or a sudden drop in ambient temperature), often only allows for passive adjustments after an error has occurred.
[0004] Following the logical deduction from existing control methods based on traditional PID error feedback, it becomes clear that due to the significant thermal inertia and large pure time lag in molten salt heating systems, when the system detects a temperature deviation and increases heating power, the heat transfer to the measuring point takes a considerable amount of time, inevitably leading to a sluggish system response. Conversely, when the temperature approaches the target value, the large amount of heat energy accumulated in the system causes severe temperature overshoot. This alternation between lag and overshoot not only causes drastic fluctuations in molten salt temperature, affecting energy storage efficiency and the safe operation of downstream equipment, but also leads to frequent switching between high power and shutdown of the heater, greatly increasing energy consumption and shortening the lifespan of heating components. Therefore, how to achieve accurate and forward-looking prediction and stable control of molten salt temperature under complex operating conditions with large time lag and nonlinearity, eliminating temperature fluctuations and overshoot, and reducing system energy consumption are the main technical problems that urgently need to be solved in current molten salt heating control technology. Summary of the Invention
[0005] To address the aforementioned issues, this application provides a machine learning-based intelligent control system for molten salt heating.
[0006] This application provides a machine learning-based intelligent control system for molten salt heating, employing the following technical solution: A machine learning-based intelligent control system for molten salt heating includes: a data acquisition module for real-time acquisition of multi-dimensional operating data of the molten salt heating system, wherein the multi-dimensional operating data includes at least molten salt temperature gradient data, flow rate data, heater power data, and ambient temperature data; a machine learning prediction module, communicatively connected to the data acquisition module, for receiving the multi-dimensional operating data and inputting it into a pre-trained time-series prediction model to output a predicted value of the molten salt temperature within a future target time period; and a heating control module, communicatively connected to the machine learning prediction module, for receiving the predicted value of the molten salt temperature, comparing the predicted value of the molten salt temperature with a target set temperature, and generating a control command based on the comparison difference to adjust the output power of the heater.
[0007] By adopting the above technical solution, during system operation, the data acquisition module continuously collects multi-dimensional operating condition data, and the machine learning prediction module uses this historical and real-time data to calculate the temperature state in the future, for example, 5 or 10 minutes later, through a time-series prediction model. The heating control module no longer relies solely on the current temperature deviation, but instead acts in advance based on the predicted temperature deviation, smoothly adjusting the heater power. The combination of feedforward prediction and feedback control effectively overcomes the thermal inertia of the molten salt system, achieving precise temperature tracking and avoiding large temperature overshoot and oscillations.
[0008] Preferably, the data acquisition module includes a temperature array of multiple thermocouple sensors disposed at different depths in the molten salt storage tank, for acquiring the molten salt temperature gradient data.
[0009] By adopting the above technical solution, the system can obtain the precise temperature of different layers inside the storage tank from bottom to top. In actual operation, the array sensors transmit the temperature of each layer to the main control unit in real time. This enables the machine learning model to perceive the thermal stratification phenomenon and convective heat transfer state inside the molten salt, thereby more accurately predicting the overall and local temperature change trends and avoiding the "blind spots" and misjudgments caused by a single temperature measurement point.
[0010] Preferably, the time-series prediction model is a deep learning model built on a Long Short-Term Memory (LSTM) network, and the loss function of the model includes a penalty term for the rate of temperature change.
[0011] By employing the above technical solution, the superior long-term dependency capture capability of LSTM is utilized to deeply mine historical data patterns of molten salt heating. The specific process is as follows: the data sequence is input into the LSTM network, which continuously updates the cell state through forget gates, input gates, and output gates; simultaneously, during backpropagation training, the loss function penalizes drastic temperature change rates. This results in a trained model with not only high prediction accuracy but also a smoother output prediction curve that conforms to the physical laws of molten salt thermodynamics.
[0012] Preferably, the machine learning prediction module further includes a feature preprocessing unit, used to denoise, normalize, and impute missing values in the collected multi-dimensional operational data, extracting temporal feature sequences which are then input into the deep learning model constructed based on the Long Short-Term Memory (LSTM) network. By employing the above technical solution, the system cleans the data before it enters the model. The operation process is as follows: first, a moving median filter is used to remove high-frequency electromagnetic interference noise from the sensor acquisition process; then, polynomial interpolation is used to fill in missing data caused by communication packet loss; finally, Min-Max normalization is performed. This significantly improves the quality of the input data, accelerates the model's inference speed, and ensures the reliability of the prediction results.
[0013] Preferably, the heating control module has a built-in adaptive fuzzy PID controller, which takes the comparison difference and the rate of change of the comparison difference as inputs and dynamically outputs the adjustment amount of the PID control parameters.
[0014] By adopting the above technical solution, the system inputs the predicted deviation and the rate of change of deviation into the fuzzy logic inference engine, and corrects the proportional, integral, and derivative coefficients online in real time according to the preset fuzzy rule table. This enables the control system to dynamically match the optimal control parameters in different heating stages (such as the rapid heating stage and the heat preservation stage), greatly enhancing the robustness and adaptability of the system.
[0015] Preferably, it also includes a safety warning module, which, based on the predicted value of the molten salt temperature output by the machine learning prediction module, triggers the action command of the physical cutting device in advance when the predicted value exceeds a safety threshold.
[0016] By adopting the above technical solution, the system achieves predictive safety protection. When the model predicts that the temperature will exceed the molten salt decomposition temperature or the equipment's tolerance limit in the future, the system will immediately issue an alarm and cut off the power supply without waiting for the actual over-temperature to occur. This avoids the dangerous situation of "temperature continuing to rise after power failure" caused by thermal inertia, greatly improving the system's safety.
[0017] Preferably, the physical disconnection device includes a solid-state relay group and an AC contactor, and the action command first controls the solid-state relay group to disconnect, and then controls the AC contactor to disconnect the main circuit.
[0018] By adopting the above technical solution, during emergency disconnection, the contactless solid-state relay first cuts off the current when the AC current crosses zero, and then disconnects the contactless AC contactor. This operation allows the contactor to operate under no-load or low-load conditions, eliminating arcing, protecting the contactor contacts, and extending the service life of the electrical hardware.
[0019] Preferably, the multi-dimensional operating data also includes the operating frequency and current data of the molten salt pump motor, and the machine learning prediction module uses the operating frequency and current data as auxiliary input features of the time series prediction model.
[0020] By adopting the above technical solution, the system incorporates the pump's operating status. During operation, the pump's frequency and current are collected to calculate the actual flow rate and circulation volume of the molten salt. Since the flow rate directly affects the convective heat transfer coefficient, it is used as an auxiliary feature input into the LSTM model, enabling the model to more accurately assess the rate at which heat is carried away, further reducing prediction errors under complex variable flow conditions.
[0021] Preferably, the heating control module is also configured with a multi-zone collaborative control strategy, which independently controls the heater output power of multiple heating zones in the molten salt storage tank according to the molten salt temperature gradient data.
[0022] By adopting the above technical solution, the system divides the storage tank into multiple heating zones, including the bottom, middle, and top. The control process is as follows: based on the gradient data fed back from the temperature array, if the bottom temperature is found to be too low, the power of the bottom heating zone is increased separately, while the power of the top zone is appropriately suppressed to prevent overheating. This effectively eliminates local temperature differences, ensures the uniformity of the temperature field inside the molten salt tank, and prevents local solidification or local overheating and degradation.
[0023] Preferably, it also includes a cloud server, which receives the multi-dimensional operational data through an industrial IoT gateway, periodically updates the parameters of the time series prediction model in the cloud using a federated learning algorithm, and then distributes the updated model to the local edge computing node.
[0024] By adopting the above technical solution, the system can achieve edge-cloud collaboration. Local nodes are responsible for low-latency real-time prediction and control, while the cloud aggregates anonymized data from multiple devices for joint training. This not only protects the data privacy of each factory area but also leverages massive samples to improve the model's generalization ability, enabling the local model to continuously evolve and adapt to characteristic drift caused by equipment aging.
[0025] In summary, this application includes at least one of the following beneficial technical effects: The system architecture includes data acquisition, machine learning prediction and heating control modules. By constructing a time-series prediction model, it can accurately predict the future temperature of molten salt and adjust the heating power based on the prediction deviation. This effectively overcomes the inherent large time delay and strong thermal inertia of molten salt heating systems, and can achieve excellent results such as stable temperature control, no large overshoot, and significant reduction in energy consumption from frequent start-stop. A model based on a long short-term memory network (LSTM) and incorporating a penalty term for the rate of temperature change is adopted. Combined with an adaptive fuzzy PID controller, the fuzzy PID controller can achieve dynamic online self-tuning of parameters due to the deep mining of nonlinear thermodynamic time series laws by LSTM, thus achieving extremely high control accuracy and robustness under complex working conditions and variable flow disturbances. Attached Figure Description
[0026] Figure 1 This is a schematic diagram of a machine learning-based intelligent control system for molten salt heating.
[0027] Figure 2 This is a schematic diagram of the internal structure of a molten salt storage tank.
[0028] Figure 3 This is a flowchart of data processing and LSTM neural network prediction.
[0029] Figure 4 This is a schematic diagram of the logic of an adaptive fuzzy PID controller.
[0030] Figure 5 This is a circuit diagram of a safety disconnection mechanism. Detailed Implementation
[0031] The present application will be further described in detail below with reference to all the accompanying drawings.
[0032] This application discloses a machine learning-based intelligent control system for molten salt heating.
[0033] Reference Figure 1 The intelligent control system for molten salt heating based on machine learning includes a data acquisition module, a machine learning prediction module, a heating control module, a safety early warning module, and a cloud server.
[0034] In practical implementation, the data acquisition module is used to collect multi-dimensional operational data of the molten salt heating system in real time. To comprehensively control the thermodynamic state inside the molten salt tank, the data acquisition module includes multiple thermocouple sensor temperature arrays installed at different depths within the molten salt tank. For example, K-type or N-type armored thermocouples (using two high-temperature resistant sensors as examples here) can be selected. They are encapsulated in high-temperature corrosion-resistant Inconel alloy sheaths and arranged at 0.5-meter intervals along the vertical direction to accurately acquire molten salt temperature gradient data. In addition, the data acquisition module also acquires flow data in the pipeline through ultrasonic or electromagnetic flow meters, reads the actual power data of the current heater through a power transmitter, and acquires ambient temperature data using an ambient temperature and humidity sensor. Simultaneously, the multi-dimensional operational data also includes the operating frequency and current data of the molten salt pump motor, which are acquired in real time through the inverter's communication interface (such as RS485 / Modbus protocol). All the above sensor nodes are connected to the data acquisition terminal in the PLC control cabinet via shielded twisted-pair cables.
[0035] The machine learning prediction module is deployed locally within an industrial-grade edge computing gateway or industrial control computer, and communicates with the data acquisition module via a high-speed industrial Ethernet. This module first includes a feature preprocessing unit to clean the acquired multi-dimensional operational data. Specifically, it uses a sliding window filtering algorithm to remove singular value noise, employs Lagrange interpolation or linear interpolation to fill in missing frame data, and uses Min-Max normalization to uniformly map all data with different dimensions to the [0, 1] interval. The preprocessed time-series feature sequences (including temperature gradient, flow rate, power, ambient temperature, pump operating frequency, and current data) are used as input to a pre-trained time-series prediction model. In this embodiment, the time-series prediction model is a deep learning model built on a Long Short-Term Memory (LSTM) network. This model includes an input layer, multiple LSTM hidden layers, and a fully connected output layer. To conform to the physical laws of molten salt heat transfer, the model's loss function incorporates a penalty term for the rate of temperature change, in addition to the traditional mean squared error (MSE). Specifically, a significant loss penalty is applied when the slope of the predicted temperature curve exceeds the physical temperature rise limit of the molten salt, ensuring that the output molten salt temperature prediction curve for the target future time period (e.g., the next 10 minutes) is smooth and consistent with reality. Since the pump's operating frequency and current data reflect the molten salt circulation rate, using them as auxiliary input features can improve the model's prediction accuracy under variable flow conditions.
[0036] The heating control module is also deployed within the local main control PLC, interacting with the machine learning prediction module in real time. After receiving the predicted molten salt temperature, the heating control module compares it with the target set temperature, calculating the prediction deviation and its rate of change. The heating control module has a built-in adaptive fuzzy PID controller. This controller has a fuzzy rule table internally, which fuzzifies the prediction deviation and its rate of change, and outputs the real-time adjustment of the PID control parameters (Kp, Ki, Kd) after fuzzy inference. Subsequently, the heating control module generates the final control command based on the adjusted parameters. Specifically, the system is configured with a multi-zone collaborative control strategy. Based on the molten salt temperature gradient data obtained from the temperature array, the storage tank is divided into multiple heating zones, such as bottom, middle, and top. The control command independently controls the output power of the heaters in multiple heating zones by adjusting the conduction angle of the thyristor voltage regulators connected to the heaters in each zone, or by adjusting the PWM duty cycle of the solid-state relays, ensuring a uniform overall temperature field.
[0037] The safety early warning module is an independent monitoring thread that continuously monitors the molten salt temperature based on the predicted value output by the machine learning prediction module. When the predicted value exceeds a preset safety threshold (e.g., 580℃, close to the critical point of molten salt decomposition), the system triggers the action command of the physical cut-off device in advance, without waiting for the actual temperature measurement point to reach the dangerous temperature. The physical cut-off device consists of a solid-state relay group and an AC contactor connected in series. The execution sequence of the action command is as follows: first, the solid-state relay group is controlled to disconnect at the zero-crossing point of the AC current to achieve arc-free and rapid current interruption; after a delay of tens of milliseconds, the AC contactor is then controlled to disconnect the main circuit to achieve physical isolation. This dual cut-off mechanism ensures both response speed and extends the contactor contact life.
[0038] The system also includes a cloud server. The cloud server receives anonymized, multi-dimensional operational data through industrial IoT gateways (supporting protocols such as MQTT) deployed on-site. To ensure data privacy and improve model generalization capabilities, the cloud regularly uses federated learning algorithms to combine device data from multiple different sites to globally update and aggregate the parameters of the time-series prediction model, and then distributes the updated model weights to each local edge computing node to achieve continuous iterative upgrades of the model.
[0039] The working principle of the machine learning-based intelligent control system for molten salt heating in this application is as follows: The data acquisition module acquires multi-dimensional holographic state data, which is preprocessed and then input into an LSTM model. The model utilizes historical memory and multivariate coupling analysis to predict the temperature trend of the system over a future period. The heating control module compares the "future temperature" with the set value and uses an adaptive fuzzy PID algorithm to provide smooth power adjustment commands in advance. Simultaneously, a multi-zone collaborative strategy addresses spatial temperature unevenness, and the safety module uses predicted data to implement proactive over-temperature protection. The cloud-based federated learning mechanism ensures the model's resistance to aging drift during long-term operation. The entire process achieves feedforward prediction in the time dimension and precise coordination in the spatial dimension, solving the control challenges caused by the large lag and high inertia of molten salt systems.
[0040] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.
Claims
1. A machine learning-based intelligent control system for molten salt heating, characterized in that, include: The data acquisition module is used to collect multi-dimensional operating data of the molten salt heating system in real time. The multi-dimensional operating data includes at least molten salt temperature gradient data, flow rate data, heater power data, and ambient temperature data. The machine learning prediction module is communicatively connected to the data acquisition module. It is used to receive the multi-dimensional operational data and input it into the pre-trained time series prediction model, and output the predicted value of molten salt temperature within the future target time period. The heating control module is communicatively connected to the machine learning prediction module. It receives the predicted value of the molten salt temperature, compares the predicted value of the molten salt temperature with the target set temperature, and generates a control command based on the comparison difference to adjust the output power of the heater.
2. The intelligent control system for molten salt heating based on machine learning according to claim 1, characterized in that, The data acquisition module includes a temperature array of multiple thermocouple sensors installed at different depths in the molten salt storage tank, used to acquire the temperature gradient data of the molten salt.
3. The intelligent control system for molten salt heating based on machine learning according to claim 1, characterized in that, The time-series prediction model is a deep learning model built on a Long Short-Term Memory (LSTM) network, and the loss function of the model includes a penalty term for the rate of temperature change.
4. The intelligent control system for molten salt heating based on machine learning according to claim 3, characterized in that, The machine learning prediction module also includes a feature preprocessing unit, which is used to denoise, normalize, and impute missing values on the collected multi-dimensional running data, and extract the temporal feature sequence to input into the deep learning model built on the Long Short-Term Memory (LSTM) network.
5. The intelligent control system for molten salt heating based on machine learning according to claim 1, characterized in that, The heating control module has a built-in adaptive fuzzy PID controller. The adaptive fuzzy PID controller takes the comparison difference and the rate of change of the comparison difference as inputs and dynamically outputs the adjustment amount of the PID control parameters.
6. The intelligent control system for molten salt heating based on machine learning according to claim 1, characterized in that, It also includes a safety warning module, which, based on the predicted molten salt temperature output by the machine learning prediction module, triggers the action command of the physical cutting device in advance when the predicted value exceeds the safety threshold.
7. The intelligent control system for molten salt heating based on machine learning according to claim 6, characterized in that, The physical disconnection device includes a solid-state relay group and an AC contactor. The action command first controls the solid-state relay group to disconnect, and then controls the AC contactor to disconnect the main circuit.
8. The intelligent control system for molten salt heating based on machine learning according to claim 1, characterized in that, The multi-dimensional operational data also includes the operating frequency and current data of the molten salt pump motor. The machine learning prediction module uses the operating frequency and current data as auxiliary input features of the time-series prediction model.
9. The intelligent control system for molten salt heating based on machine learning according to claim 1, characterized in that, The heating control module is also equipped with a multi-zone collaborative control strategy, which independently controls the output power of heaters in multiple heating zones within the molten salt storage tank based on the molten salt temperature gradient data.
10. The intelligent control system for molten salt heating based on machine learning according to claim 1, characterized in that, It also includes a cloud server, which receives the multi-dimensional operational data through an industrial IoT gateway, periodically updates the parameters of the time series prediction model in the cloud using a federated learning algorithm, and then distributes the updated model to the local edge computing node.