Thermal power generating unit frequency modulation control method based on deep learning

Abstract

The invention relates to the technical field of thermal power generating unit frequency modulation, and discloses a thermal power generating unit frequency modulation control method based on deep learning. According to the method, through real-time acquisition and preprocessing of operation data, power grid data, environment data and historical frequency modulation data of a thermal power generating unit, abnormal values and missing values can be effectively removed, amplitude and periodic characteristics of frequency fluctuation are extracted, frequency prediction is carried out by adopting a long-short-term memory network model, and the accuracy of frequency prediction is improved. According to the method, complex time sequence characteristics and dynamic changes can be captured, the prediction precision is remarkably improved, compared with a traditional control method depending on experience, the method has higher adaptability and flexibility, in addition, the needed adjusting amount can be accurately calculated in combination with LSTM output and a PID control strategy, and therefore rapid and accurate frequency adjustment is achieved, and the method is suitable for large-scale popularization and application. The response speed and the adjusting capacity of the thermal power generating unit are improved, the stability and the reliability of the power system are enhanced, and finally efficient and safe power supply is achieved.

Description

Technical Field

[0001] This invention relates to the field of frequency regulation technology for thermal power units, specifically a frequency regulation control method for thermal power units based on deep learning. Background Technology

[0002] Thermal power units, as a core component of the power system, bear the responsibility of providing a stable and reliable power supply to the grid. Their operational stability and reliability are crucial to ensuring the normal operation of the power system. However, the rapid development of renewable energy sources (such as wind and solar power) in recent years has brought profound changes to the power system. The power generation characteristics of renewable energy are highly volatile and uncertain, especially under the influence of factors such as climate change and day-night cycles, posing unprecedented challenges to the stability of grid frequency.

[0003] Traditional frequency regulation control methods mainly rely on human experience and simple control algorithms. These methods can meet basic frequency regulation requirements in most cases, but they often fail to achieve efficient frequency regulation when faced with complex and dynamic power system environments. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning-based frequency regulation control method for thermal power units. By real-time acquisition and preprocessing of operating data, grid data, environmental data, and historical frequency regulation data of thermal power units, it can effectively remove outliers and missing values ​​and extract the amplitude and periodic characteristics of frequency fluctuations, thus providing a high-quality data foundation for subsequent analysis. Secondly, the use of a long short-term memory network model for frequency prediction can capture complex time-series characteristics and dynamic changes, significantly improving prediction accuracy. Compared with traditional experience-based control methods, it has stronger adaptability and flexibility. Furthermore, by combining LSTM output with a PID control strategy, the required adjustment amount can be accurately calculated, thereby achieving fast and accurate frequency regulation. This intelligent frequency regulation control strategy not only improves the response speed and regulation capability of thermal power units but also enhances the stability and reliability of the power system, providing strong support for coping with the volatility of renewable energy, and ultimately achieving efficient and secure power supply.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution: a frequency regulation control method for thermal power units based on deep learning, comprising the following steps:

[0008] S1. Real-time acquisition of operating data of thermal power units, power grid data, environmental data, and historical frequency regulation data;

[0009] S2. Preprocess the collected operating data, power grid data, environmental data and historical frequency regulation data of thermal power units, remove outliers and missing values ​​and normalize them, and extract the amplitude characteristics and periodic characteristics of frequency fluctuations.

[0010] S3. Select a long short-term memory network model. Input the preprocessed thermal power unit operation data, power grid data, environmental data, and historical frequency regulation data, along with the extracted amplitude and periodic characteristics of frequency fluctuations, into the long short-term memory network model. Use a loss function for training and perform cross-validation to adjust the model hyperparameters.

[0011] S4. Monitor and collect operating data of thermal power units and power grid data in real time, and input the operating data of thermal power units and power grid data into the trained long short-term memory network model for frequency prediction;

[0012] S5. Based on the frequency fluctuations predicted by the long short-term memory network model, set the frequency modulation target, calculate the required adjustment amount using PID control based on the frequency prediction results, and generate a frequency modulation control strategy by combining the output of the long short-term memory network model.

[0013] S6. The control signal generated according to the frequency modulation control strategy is sent to the actual control system to implement the frequency modulation control strategy.

[0014] Preferably, the formula for calculating the amplitude characteristics of the frequency fluctuation is as follows:

[0015]

[0016] In the formula, Δf amp f represents the amplitude characteristic of frequency fluctuations. i This represents the frequency value of the i-th sampling point. This represents the average frequency within the time window, where N represents the number of sampling points within the time window, and i represents the index.

[0017] Preferably, the formula for calculating the periodic characteristics of the frequency fluctuation is as follows:

[0018]

[0019] In the formula, r k The periodic characteristic of frequency fluctuation, f i This represents the i-th frequency sample value. The frequency average value within the time window is represented by k, the lag step size is represented by the time delay when calculating the autocorrelation, N represents the number of sampling points within the time window, and i represents the index subscript.

[0020] Preferably, the Long Short-Term Memory (LSTM) network model needs to update its hidden states constantly, and the update formula is as follows:

[0021] h t =LSTM(x t ,h t-1 )

[0022] In the formula, h t Let x represent the hidden state at time step t. LSTM(*) represents the internal operation of the Long Short-Term Memory network, which updates the hidden state using a gating mechanism. t This represents the input feature vector at time step t, which includes the preprocessed raw data and the extracted features.

[0023] Preferably, the loss function is as follows:

[0024]

[0025] In the formula, L represents the output value of the loss function. This represents the future frequency value predicted by the model for the i-th sample. Let M represent the actual observation frequency of the i-th sample, and M represent the batch size.

[0026] Preferably, the frequency prediction may have a deviation, which is calculated using the following formula:

[0027]

[0028] In the formula, Indicates prediction bias, used to assess the degree to which future frequencies deviate from the target, f target Indicates the target frequency of the system. This represents the future frequency values ​​predicted by the model.

[0029] Preferably, the formula used for calculating the adjustment amount required by the PID control is as follows:

[0030]

[0031] In the formula, u(t) represents the adjustment control output, e(t) represents the current frequency deviation, and K... p K represents the proportionality coefficient, which controls the current error magnitude. i K represents the integral coefficient used to eliminate steady-state deviation. d These represent the differential coefficients, used to predict changes in error. t represents the integral of the current frequency deviation, and t represents time.

[0032] Preferably, the frequency modulation control strategy is implemented using the following formula:

[0033]

[0034] In the formula, u just This represents the command generated by the frequency modulation control strategy, α represents the adjustment parameter that determines the weights of PID control and model predictive regulation, and u(t) represents the PID control output. This indicates prediction bias.

[0035] Preferably, the control signal is further optimized using a fuzzy logic system.

[0036] Preferably, the control strategy incorporates the following fuzzy rules:

[0037]

[0038] The fuzzy adjustment, as the final control signal, is used to regulate the output of the thermal power unit. Here, "fuzzy" represents the fuzzy inference function, which applies a set of fuzzy rules based on the input parameters to output an adjustment amount. u(Δt) represents the adjustment amount at the current time t. This indicates prediction bias.

[0039] Compared with existing technologies, this invention provides a deep learning-based frequency regulation control method for thermal power units, which has the following advantages:

[0040] This invention effectively removes outliers and missing values ​​and extracts the amplitude and periodic characteristics of frequency fluctuations by real-time acquisition and preprocessing of operating data of thermal power units, power grid data, environmental data, and historical frequency regulation data. This provides a high-quality data foundation for subsequent analysis. Secondly, it uses a long short-term memory network model for frequency prediction, which can capture complex time-series characteristics and dynamic changes, significantly improving prediction accuracy. Compared with traditional experience-based control methods, it has stronger adaptability and flexibility. In addition, by combining LSTM output with PID control strategy, it can accurately calculate the required adjustment amount, thereby achieving fast and accurate frequency regulation. This intelligent frequency regulation control strategy not only improves the response speed and regulation capability of thermal power units, but also enhances the stability and reliability of the power system, providing strong support for coping with the volatility of renewable energy, and ultimately achieving efficient and safe power supply. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Traditional frequency regulation control methods rely primarily on human experience and simple control algorithms. While these methods can meet basic frequency regulation requirements in most cases, they often fail to achieve efficient frequency regulation in complex and dynamic power system environments. Therefore, a deep learning-based frequency regulation control method for thermal power units is proposed. (See [link to relevant documentation]). Figure 1 The method includes the following steps:

[0044] S1. Real-time acquisition of operating data of thermal power units, power grid data, environmental data, and historical frequency regulation data;

[0045] In practical applications, to ensure the accuracy and timeliness of frequency regulation control for thermal power units, multiple types of sensors and data acquisition equipment are used to monitor the operating status of the units in real time, including key operating parameters such as generator output power, speed, turbine parameters, fuel supply status, and mechanical vibration. Simultaneously, a power grid monitoring system is used to acquire real-time electrical parameters such as grid frequency, voltage, current, and power factor, enabling precise control of grid load changes and frequency fluctuations. Regarding environmental data, temperature, pressure, and humidity sensors are deployed to collect real-time ambient temperature, pressure, humidity, and other meteorological parameters, ensuring that the regulation strategy considers the impact of environmental factors on unit performance. In addition to the impact of energy, a historical frequency modulation database is constructed by utilizing the frequency modulation command records, frequency modulation response measurement results, and historical frequency deviation trend data of historical frequency modulation equipment. This data is collected in real time through high-speed data acquisition networks (such as industrial Ethernet, fieldbus, etc.) and transmitted to a remote data processing platform. After standardization, time synchronization, and preprocessing, a rich multi-source information foundation is formed, which provides strong technical support for subsequent deep learning model training, frequency prediction, and frequency modulation strategy formulation. This series of technical means ensures the real-time performance, accuracy, and integrity of the data, laying a solid foundation for the efficient and safe operation of the intelligent frequency modulation control system.

[0046] S2. Preprocess the collected operating data, power grid data, environmental data and historical frequency regulation data of thermal power units, remove outliers and missing values ​​and normalize them, and extract the amplitude characteristics and periodic characteristics of frequency fluctuations.

[0047] In the process of preprocessing the collected thermal power unit operation data, power grid data, environmental data, and historical frequency regulation data, firstly, anomaly detection algorithms (such as Z-score method, box plot method, or deep learning-based anomaly detection model) are used to automatically identify and remove outliers in the data to ensure that extreme values ​​will not interfere with subsequent modeling. Secondly, for the problem of missing values, interpolation completion techniques (such as linear interpolation, polynomial interpolation, or K-nearest neighbor-based interpolation) are applied to supplement the missing parts of the data to ensure the continuity of time series data. Data preprocessing also includes normalization of all data. Common methods such as extremum standardization or z-score standardization are used to map feature data of different dimensions to a unified range (such as [0,1] or mean of 0 and variance of 1) to improve the training efficiency and prediction stability of deep learning models.

[0048] In the feature extraction stage, to accurately describe the changing state of frequency deviation, the amplitude and periodic characteristics of frequency fluctuations are calculated. The amplitude of frequency fluctuations is calculated using the following formula:

[0049]

[0050] Among them, i i This represents the frequency value of the i-th sampling point. This formula represents the average frequency over the sampling period. By calculating the standard deviation of the frequency sequence, it quantifies the magnitude of frequency fluctuations and provides important deviation information for frequency modulation strategies.

[0051] Furthermore, in order to identify the periodic characteristics of frequency signals, the autocorrelation coefficient is used to measure the similarity of signals under different hysteresis k values, as shown in the following formula:

[0052]

[0053] Among them, f i For the i-th frequency sampling point, Here, N is the average frequency, N is the total number of sampling points, and k is the lag coefficient. This formula reveals the periodic structure of the frequency signal by calculating the normalized autocorrelation coefficient under different lags k, and identifies representative periodic features. For example, when a certain k value corresponds to a high autocorrelation coefficient, it means that the corresponding frequency oscillation has obvious periodic features, providing a strong basis for predicting frequency fluctuations.

[0054] S3. Select a long short-term memory network model. Input the preprocessed thermal power unit operation data, power grid data, environmental data, and historical frequency regulation data, along with the extracted amplitude and periodic characteristics of frequency fluctuations, into the long short-term memory network model. Use a loss function for training and perform cross-validation to adjust the model hyperparameters.

[0055] In the stage of establishing a deep learning model for the frequency regulation system of thermal power units, a Long Short-Term Memory (LSTM) network is used as the core prediction model to fully capture the temporal characteristics of frequency changes and achieve accurate frequency deviation prediction. First, after preprocessing the operating data of thermal power units, power grid electrical parameters, environmental data and historical frequency regulation data, the amplitude characteristics (such as standard deviation) and periodic characteristics (such as autocorrelation coefficient) of frequency fluctuations are extracted. These characteristics are used as the input vector of the LSTM model, combining static features with dynamic time series information to provide rich prediction information for the model.

[0056] During model training, the hidden state at each time step is implemented using the following recursive formula:

[0057] h t =LSTM(x t ,h t-1 )

[0058] Here, x t This represents the input feature vector at the current time step t, containing all preprocessed key metrics; h t-1 The hidden state representing the previous time step reflects the ability to remember previous information. LSTM utilizes its gating mechanism to effectively alleviate the gradient vanishing problem of traditional RNN in long sequences and ensure long-distance dependency capture of historical information.

[0059] To optimize model performance, a suitable loss function (such as mean squared error, MSE) is used for training:

[0060]

[0061] in, Let be the model-predicted frequency value of the i-th sample at time t+1. M represents the corresponding actual observed value, and M is the batch sample size. This formula improves prediction accuracy by gradually adjusting model parameters by minimizing the mean square error between the predicted and actual values. During training, cross-validation and other validation methods are used to dynamically adjust model hyperparameters (such as the number of hidden layer units, learning rate, sequence length, regularization parameters, etc.) to ensure the model's generalization ability on the training and validation sets and avoid overfitting. This results in the construction of a frequency prediction model with good robustness, providing an effective basis for subsequent frequency modulation control.

[0062] S4. Monitor and collect operating data of thermal power units and power grid data in real time, and input the operating data of thermal power units and power grid data into the trained long short-term memory network model for frequency prediction;

[0063] In the operation of the frequency regulation control system of thermal power units, the system monitors and collects the working status parameters of thermal power units and key electrical indicators of the power grid such as frequency, voltage, and current in real time to ensure the timeliness and accuracy of the data. The collected thermal power unit operating data includes dynamic parameters such as generator output power, speed, pressure, and temperature, while the power grid data includes system frequency, voltage deviation, and load changes. These multi-source real-time data are transmitted to the data processing platform through high-speed data acquisition devices and industrial communication protocols (such as fieldbus, Ethernet, etc.). After preprocessing (such as outlier removal, missing value imputation, normalization, etc.), an input vector is formed, which serves as the input feature of the deep learning model (LSTM) for frequency deviation prediction.

[0064] The trained Long Short-Term Memory (LSTM) network model processes these input features and outputs frequency predictions for future time points. In actual monitoring, the latest input data (thermal power unit parameters and grid frequency indicators) are fed into the model in real time to obtain frequency predictions. Based on the model's output, the predicted value of future frequency deviation is calculated, as shown in the following formula:

[0065]

[0066] f here target This indicates the preset target frequency of the power grid (e.g., 50Hz or 60Hz), representing the ideal frequency value of the system. It represents the frequency value of future time points predicted by the deep learning model, reflecting the trend of the system deviating from the target frequency in the future;

[0067] The essence of this formula is to intuitively express the future frequency deviation of the system by predicting the difference between the future frequency value and the target frequency, thus providing a key basis for formulating the next frequency modulation strategy. (Frequency deviation) It can be used as the core parameter for regulation and control. Combined with regulation strategies (such as PID control) or fuzzy logic optimization methods, it can dynamically adjust the output power of thermal power units to ensure the stability of system frequency and achieve the goal of efficient and intelligent frequency regulation. The whole process relies on high-speed and continuous monitoring, data transmission, and model inference, which reflects the core means of advanced data-driven frequency regulation control technology.

[0068] S5. Based on the frequency fluctuations predicted by the long short-term memory network model, set the frequency modulation target, calculate the required adjustment amount using PID control based on the frequency prediction results, and generate a frequency modulation control strategy by combining the output of the long short-term memory network model.

[0069] In the frequency regulation system of thermal power units, leveraging the predictive capabilities of the Long Short-Term Memory (LSTM) network model, the target frequency for future frequency regulation is first set based on the predicted frequency deviation to ensure frequency stability within the expected range. Then, the current frequency deviation error value is calculated using the prediction results. This error value reflects the difference between the predicted frequency and the target frequency, providing fundamental data for the regulation strategy. Based on this error, the required regulation amount is calculated using the classic Proportional-Integral-Derivative (PID) control algorithm, as shown in the following formula:

[0070]

[0071] Among them, K p K represents the proportional gain, used to adjust the contribution of the current error to the control output, and is a fast-response frequency deviation. i K represents the integral coefficient, used to eliminate long-term persistent deviations and ensure that the frequency remains stable near the target value. d These represent the differential coefficients, used to predict error trends, mitigate system overshoot, and improve dynamic response.

[0072] In practical applications, the dynamic change of error e(t) is accumulated through integration, representing the cumulative deviation and responding to the persistent deviation of the system; the differential term, through the calculation of the error change rate, intervenes and adjusts in advance, enhancing the stability and response speed of the system. The calculated adjustment amount u(t) is combined with the output of the deep learning model to form a complete frequency modulation control strategy, with the frequency deviation predicted by the model guiding the future frequency change trend.

[0073] PID control determines the instantaneous control signal: adjusts the generating power of the thermal power unit to correct frequency deviation;

[0074] This system achieves a perfect integration of data-driven and classical control methods, improves the intelligence and response efficiency of frequency regulation control, and ensures the stable operation of the power grid frequency.

[0075] S6. The control signal generated according to the frequency modulation control strategy is sent to the actual control system to implement the frequency modulation control strategy.

[0076] In implementing the frequency modulation control strategy, the first step is to compare the adjustment signal u(t) calculated by the PID controller with the future frequency predicted by the deep learning model. Combined, they form a fusion regulation quantity u just To enhance the adaptability and stability of the frequency modulation system, the formula for calculating the fusion adjustment amount is as follows:

[0077]

[0078] The parameter α (ranging from 0 to 1) is used to adjust the weights of traditional PID control and deep learning prediction, thereby combining the advantages of both. To further improve the robustness and response speed of frequency modulation control, a fuzzy logic system is also introduced, which uses fuzzy rules to optimize the control signal. Specifically, the fuzzy system optimizes the control signal based on the current control signal u(t) and the predicted frequency deviation. Output a fuzzy adjustment amount:

[0079]

[0080] This fuzzy adjustment is achieved through a series of predefined fuzzy rules, such as "if u(Δr) is large and..." "If it is high, reduce the adjustment intensity", or "If u(Δr) is small and..." "If the frequency is too low, the adjustment intensity will be increased." Based on a fuzzy rule base and combined with a fuzzy inference mechanism, the control signal is adjusted in real time. Finally, the fuzzy-optimized adjustment signal is used as the control signal and transmitted to the actual control system of the thermal power unit to adjust the output power of the generator unit and achieve the stability of the target frequency. This fusion control strategy combines the fast response of traditional PID control, the long-term prediction capability of deep learning models, and the robustness of fuzzy logic systems. It effectively enhances the adaptability and control accuracy of the frequency regulation system in complex environments and provides a strong technical guarantee for the stability of the power grid frequency.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A frequency regulation control method for thermal power units based on deep learning, characterized in that, Includes the following steps: S1. Real-time acquisition of operating data of thermal power units, power grid data, environmental data, and historical frequency regulation data; S2. Preprocess the collected operating data, power grid data, environmental data and historical frequency regulation data of thermal power units, remove outliers and missing values ​​and normalize them, and extract the amplitude characteristics and periodic characteristics of frequency fluctuations. S3. Select a long short-term memory network model. Input the preprocessed thermal power unit operation data, power grid data, environmental data, and historical frequency regulation data, along with the extracted amplitude and periodic characteristics of frequency fluctuations, into the long short-term memory network model. Use a loss function for training and perform cross-validation to adjust the model hyperparameters. S4. Monitor and collect operating data of thermal power units and power grid data in real time, and input the operating data of thermal power units and power grid data into the trained long short-term memory network model for frequency prediction; S5. Based on the frequency fluctuations predicted by the long short-term memory network model, set the frequency modulation target, calculate the required adjustment amount using PID control based on the frequency prediction results, and generate a frequency modulation control strategy by combining the output of the long short-term memory network model. S6. The control signal generated according to the frequency modulation control strategy is sent to the actual control system to implement the frequency modulation control strategy.

2. The frequency regulation control method for thermal power units based on deep learning according to claim 1, characterized in that: The formula for calculating the amplitude characteristics of the frequency fluctuation is as follows: In the formula, Δf amp f represents the amplitude characteristic of frequency fluctuations. i This represents the frequency value of the i-th sampling point. This represents the average frequency within the time window, where N represents the number of sampling points within the time window, and i represents the index.

3. The frequency regulation control method for thermal power units based on deep learning according to claim 2, characterized in that: The formula for calculating the periodic characteristics of the frequency fluctuation is as follows: In the formula, r k The periodic characteristic of frequency fluctuation, f i This represents the i-th frequency sample value. The frequency average value within the time window is represented by k, the lag step size is represented by the time delay when calculating the autocorrelation, N represents the number of sampling points within the time window, and i represents the index subscript.

4. The frequency regulation control method for thermal power units based on deep learning according to claim 3, characterized in that: The Long Short-Term Memory (LSTM) network model requires constant updates to its hidden states, using the following formula: h t =LSTM(x t ,h t-1 ) In the formula, h t Let x represent the hidden state at time step t. LSTM(*) represents the internal operation of the Long Short-Term Memory network, which updates the hidden state using a gating mechanism. t This represents the input feature vector at time step t, which includes the preprocessed raw data and the extracted features.

5. The frequency regulation control method for thermal power units based on deep learning according to claim 4, characterized in that: The loss function is as follows: In the formula, L represents the output value of the loss function. This represents the future frequency value predicted by the model for the i-th sample. Let M represent the actual observation frequency of the i-th sample, and M represent the batch size.

6. The frequency regulation control method for thermal power units based on deep learning according to claim 5, characterized in that: The frequency prediction may have a deviation, which is calculated using the following formula: In the formula, Indicates prediction bias, used to assess the degree to which future frequencies deviate from the target, f target Indicates the target frequency of the system. This represents the future frequency values ​​predicted by the model.

7. The frequency regulation control method for thermal power units based on deep learning according to claim 6, characterized in that: The formula used to calculate the required adjustment amount for the PID control is as follows: In the formula, u(t) represents the adjustment control output, e(t) represents the current frequency deviation, and K... p K represents the proportionality coefficient, which controls the current error magnitude. i K represents the integral coefficient used to eliminate steady-state deviation. d These represent the differential coefficients, used to predict changes in error. t represents the integral of the current frequency deviation, and t represents time.

8. The frequency regulation control method for thermal power units based on deep learning according to claim 7, characterized in that: The frequency modulation control strategy is implemented through the following formula: In the formula, u just This represents the command generated by the frequency modulation control strategy, α represents the adjustment parameter that determines the weights of PID control and model predictive regulation, and u(t) represents the PID control output. This indicates prediction bias.

9. The frequency regulation control method for thermal power units based on deep learning according to claim 8, characterized in that: The control signal is also optimized using a fuzzy logic system.

10. The frequency regulation control method for thermal power units based on deep learning according to claim 9, characterized in that: The control strategy incorporates the following fuzzy rules: The fuzzy adjustment, as the final control signal, is used to regulate the output of the thermal power unit. Here, "fuzzy" represents the fuzzy inference function, which applies a set of fuzzy rules based on the input parameters to output an adjustment amount. u(Δt) represents the adjustment amount at the current time t. This indicates prediction bias.

Images