Frequency modulation instruction prediction method and system for super capacitor coupling lithium battery energy storage

Abstract

The invention relates to the technical field of instruction prediction, and discloses a frequency modulation instruction prediction method and system for super capacitor coupling lithium battery energy storage. The method comprises the following steps: performing decomposition processing on a standardized frequency modulation sequence through a smoothing function to obtain an approximate integratable amount and an inintegratable amount; performing trigonometric function periodic conversion processing on the approximate integrable quantity to obtain a periodic characteristic sequence which is composed of a plurality of sine and cosine functions with different frequencies; carrying out linear merging transformation processing on the periodic feature sequence and the non-integrable quantity to obtain a combined feature sequence; performing training processing on the combined feature sequence through a deep neural network to obtain a frequency modulation instruction prediction model; and performing prediction processing on the real-time frequency modulation instruction data through the frequency modulation instruction prediction model to obtain a prediction result, and performing reverse normalization processing on the prediction result to obtain a frequency modulation instruction prediction value. According to the invention, the efficiency and accuracy of frequency modulation instruction prediction for super capacitor coupling lithium battery energy storage are improved.

Description

Technical Field

[0001] The present application relates to the field of instruction prediction, and in particular to a method and system for predicting frequency modulation instructions for supercapacitor-coupled lithium battery energy storage. Background Art

[0002] In the power system, thermal power units are the main frequency regulation power source and undertake the important task of power grid frequency regulation. With the large-scale grid connection of new energy and the improvement of environmental protection requirements, the power grid's demand for frequency regulation services continues to increase. When performing frequency regulation, traditional thermal power units need to frequently change the unit output. This mode of operation will lead to problems such as increased coal consumption, increased equipment wear, and shortened operating life. To solve this problem, introducing energy storage systems as auxiliary frequency regulation devices into thermal power plants has become an important technical solution. Among them, the hybrid energy storage system composed of supercapacitors and lithium batteries is widely used in the field of power frequency regulation due to its advantages such as high power density, high energy density, and fast response speed.

[0003] However, the existing energy storage system auxiliary frequency regulation technology has the problem of insufficient prediction accuracy. Since the frequency regulation instructions are nonlinear and irregular, traditional prediction methods often directly input the original frequency regulation instructions into the prediction model without fully considering the mathematical characteristics of the signal, resulting in large errors in the prediction results. This lack of prediction accuracy will affect the power allocation decision of the energy storage system, reduce the frequency regulation performance of the system, and fail to give full play to the advantages of the hybrid energy storage system. Summary of the invention

[0004] The present application provides a method and system for predicting frequency modulation instructions for supercapacitor-coupled lithium battery energy storage, which are used to improve the efficiency and accuracy of frequency modulation instruction prediction for supercapacitor-coupled lithium battery energy storage.

[0005] In a first aspect, the present application provides a method for predicting frequency modulation instructions for supercapacitor-coupled lithium battery energy storage, the method for predicting frequency modulation instructions for supercapacitor-coupled lithium battery energy storage comprising: performing data cleaning and normalization processing on an original frequency modulation instruction sequence to obtain a standardized frequency modulation sequence; decomposing the standardized frequency modulation sequence through a smoothing function to obtain an approximate integrable quantity and an integrable quantity; periodically transforming the approximate integrable quantity through a trigonometric function to obtain a periodic feature sequence, wherein the periodic feature sequence is composed of a plurality of sine and cosine functions of different frequencies; performing linear merging transformation processing on the periodic feature sequence and the integrable quantity through a weight coefficient to obtain a combined feature sequence; training the combined feature sequence through a deep neural network to obtain a frequency modulation instruction prediction model, the frequency modulation instruction prediction model comprising a three-layer convolutional neural network, a bidirectional long short-term memory network and a fully connected layer; predicting real-time frequency modulation instruction data through the frequency modulation instruction prediction model to obtain a prediction result, and denormalizing the prediction result to obtain a frequency modulation instruction prediction value.

[0006] In a second aspect, the present application provides a frequency modulation instruction prediction system for supercapacitor coupled lithium battery energy storage, the frequency modulation instruction prediction system for supercapacitor coupled lithium battery energy storage comprising: A processing module is used to perform data cleaning and normalization processing on the original frequency modulation instruction sequence to obtain a standardized frequency modulation sequence; A decomposition module, used for decomposing the standardized frequency modulation sequence through a smoothing function to obtain an approximate integrable quantity and an integrable quantity; A conversion module, used for performing periodic conversion processing on the approximate integrable quantity through trigonometric functions to obtain a periodic characteristic sequence, wherein the periodic characteristic sequence is composed of a plurality of sine and cosine functions of different frequencies; A transformation module, used for performing a linear merging transformation process on the periodic feature sequence and the non-integrable quantity through a weight coefficient to obtain a combined feature sequence; A training module, used for training the combined feature sequence through a deep neural network to obtain a frequency modulation instruction prediction model, wherein the frequency modulation instruction prediction model includes a three-layer convolutional neural network, a bidirectional long short-term memory network and a fully connected layer; The prediction module is used to perform prediction processing on the real-time frequency modulation instruction data through the frequency modulation instruction prediction model to obtain a prediction result, and perform denormalization processing on the prediction result to obtain a frequency modulation instruction prediction value.

[0007] In the technical solution provided by the present application, by performing data cleaning and normalization on the original frequency modulation instruction sequence, the missing values, outliers and noise interference in the data are effectively removed, the reliability of the data quality is ensured, and the standardized frequency modulation sequence is decomposed by a smoothing function, and the complex nonlinear sequence is decomposed into two parts: an approximately integrable quantity and an integrable quantity, which is convenient for separate processing and improves the prediction accuracy. The approximate integrable quantity is subjected to trigonometric function periodic conversion processing, and the irregular time series is converted into a periodic function combination with clear mathematical characteristics, which enhances the predictability of the signal. The periodic feature sequence and the integrable quantity are subjected to linear merging transformation processing by weight coefficients, which realizes the organic combination of periodic changes and random fluctuations, so that the prediction model can capture the regularity and random characteristics of the signal at the same time. The combined feature sequence is trained and processed by a deep neural network, which makes full use of the advantages of the convolutional neural network in extracting local features and the long short-term memory network in capturing long-term dependencies, and a model with strong prediction ability is constructed. In the real-time prediction stage, the accuracy and practicality of the prediction results are ensured by preprocessing and anti-normalization. The overall solution significantly improves the prediction accuracy of frequency modulation instructions through the system design of data preprocessing, feature extraction, model training and prediction output. Experimental verification shows that the prediction RMSE of frequency modulation instructions 1 and 2 is 1.2 and 3.3 respectively, which is significantly improved compared with 9.9 and 8.7 of the traditional model, providing a reliable decision-making basis for the efficient operation of supercapacitor and lithium battery hybrid energy storage system. This improvement in prediction accuracy helps to optimize the power allocation strategy of the energy storage system, improve the frequency modulation performance of the system, extend the service life of the equipment, and reduce operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0009] Figure 1 A schematic diagram of an embodiment of a method for predicting frequency modulation instructions for supercapacitor-coupled lithium battery energy storage in an embodiment of the present application; Figure 2 Schematic diagram of an embodiment of a frequency modulation instruction prediction system for supercapacitor coupled lithium battery energy storage in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The embodiment of the present application provides a method and system for predicting frequency modulation instructions for supercapacitor coupled lithium battery energy storage. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0011] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage in the embodiment of the present application includes: Step S101, performing data cleaning and normalization processing on the original frequency modulation instruction sequence to obtain a standardized frequency modulation sequence; Step S102, decomposing the standardized frequency modulation sequence by a smoothing function to obtain approximate integrable quantities and integrable quantities; Step S103, performing periodic conversion processing on the approximate integrable quantity through trigonometric functions to obtain a periodic characteristic sequence, wherein the periodic characteristic sequence is composed of a plurality of sine and cosine functions of different frequencies; Step S104, performing linear merging transformation processing on the periodic feature sequence and the integrable quantity through the weight coefficient to obtain a combined feature sequence; Step S105, training the combined feature sequence through a deep neural network to obtain a frequency modulation instruction prediction model, where the frequency modulation instruction prediction model includes a three-layer convolutional neural network, a bidirectional long short-term memory network, and a fully connected layer; Step S106: The real-time frequency modulation instruction data is predicted by the frequency modulation instruction prediction model to obtain a prediction result, and the prediction result is denormalized to obtain a frequency modulation instruction prediction value.

[0012] It is understandable that the execution subject of the present application can be a frequency modulation instruction prediction system for supercapacitor coupled lithium battery energy storage, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0013] Specifically, the original frequency modulation instruction sequence is cleaned and normalized. The process includes key steps such as missing value detection, outlier detection, time window division, amplitude normalization and frequency domain filtering. In the missing value detection link, by scanning the data points in the original frequency modulation instruction sequence, the positions where the values ​​are empty or unreasonable are marked to obtain the missing value position information. The linear interpolation method is used to repair the data at these positions to achieve the integrity of the sequence. Then, outlier detection is performed, and statistical methods are used to identify data points that deviate significantly from the overall distribution, mark them as outliers, and replace them with the median to ensure the rationality of the data. Then the complete sequence is divided into time windows, and a fixed time length is set, such as 5 minutes as a window, and a 75% overlap rate is used for sliding to increase the relevance of the data. For the divided sequence samples, the amplitude range is calculated, min-max normalization is performed, and the values ​​are mapped to the 0-1 interval. Finally, the frequency characteristics of the sequence are analyzed by Fourier transform, and a bandpass filter is designed to remove high-frequency noise to obtain a standardized frequency modulation sequence. After obtaining the standardized frequency modulation sequence, a smoothing function is used to decompose it, with the aim of dividing the sequence into two parts: integrable and non-integrable. Specifically, the Gaussian kernel function is used as a smoothing tool, and the smoothing degree of the sequence is controlled by adjusting the window length parameter. The parameters of the Gaussian kernel function are adaptively determined by the volatility characteristics of the sequence to ensure the rationality of the smoothing effect. The smoothed sequence is used as an approximate integrable quantity, and the difference between the original sequence and the approximate integrable quantity constitutes a non-integrable quantity. This decomposition method effectively distinguishes the main trend and random volatility components of the sequence.

[0014] For the approximate integrable quantity obtained by decomposition, it is necessary to perform trigonometric function periodic transformation. First, estimate the fundamental period of the sequence, and use the autocorrelation analysis method to find the main periodic characteristics of the signal. Then, the sequence is segmented according to the periodic parameters to facilitate subsequent harmonic analysis. Harmonic decomposition is performed on the segmented sequence to extract the fundamental wave and each harmonic component. These components are fitted using sine and cosine functions to obtain the corresponding amplitude and phase coefficients. Finally, each frequency component is superimposed according to the size of the coefficient to reconstruct the periodic characteristic sequence. In order to comprehensively utilize the information of the periodic characteristic sequence and the integrable quantity, a linear merging transformation is required. First, the importance of the periodic characteristic sequence is evaluated, and its contribution to the original sequence is calculated to obtain the corresponding weight coefficient. At the same time, the fluctuation characteristics of the integrable quantity are analyzed, and the weight coefficient is determined according to its statistical characteristics. The two parts of the sequence are multiplied by the corresponding weight coefficients and linearly combined to obtain a combined sequence that retains both periodicity and random fluctuation characteristics.

[0015] The processing of the combined feature sequence is trained using a deep neural network, and the network structure includes a convolutional layer, a recurrent layer, and a fully connected layer. First, the sequence is divided into a training set and a validation set, and an appropriate batch size is set for data loading. The local features are extracted through a three-layer convolutional neural network, and the convolution kernel size and number of channels of each layer are set according to the data characteristics. Then a bidirectional long short-term memory network is used to process the temporal dependency and capture the long-term and short-term change patterns. Finally, the features are mapped to the prediction space through a fully connected layer to obtain a prediction model.

[0016] Preprocess the frequency modulation instruction data collected in real time to make its format and features consistent with the training data. Input the preprocessed data into the prediction model and obtain the preliminary prediction value through forward calculation. Perform validity constraint check on the prediction value to ensure that it meets the physical constraint conditions. Finally, restore the prediction result to the original value range through inverse normalization transformation to obtain the final frequency modulation instruction prediction value.

[0017] Taking an actual thermal power plant frequency regulation scenario as an example, in the original frequency regulation instruction sequence collected by a thermal power unit within one hour, one data point is recorded every 5 seconds, totaling 720 data points. Through data cleaning, it is found that there are 8 missing values ​​and 12 outliers, which are corrected by linear interpolation and median replacement respectively. The data is segmented with a window length of 5 minutes and a sliding step of 75 seconds to obtain 48 sample sequences. These sequences are normalized, and the original amplitude range of -50MW to +50MW is mapped to the 0-1 interval. The approximate integrable quantity obtained by smoothing decomposition shows obvious 15-minute periodic characteristics, and the standard deviation of the integrable quantity is 0.15. After periodic conversion, three main frequency components are obtained, with amplitudes of 0.8, 0.5 and 0.3 respectively. After weight optimization, the combined weight ratio of periodic feature sequence and integrable quantity is 7:3. The root mean square error of the final model on the test set is 1.8MW, which improves the prediction accuracy by 65% ​​compared with the traditional method.

[0018] In the embodiment of the present application, by performing data cleaning and normalization on the original frequency modulation instruction sequence, the missing values, outliers and noise interference in the data are effectively removed, the reliability of the data quality is ensured, and the standardized frequency modulation sequence is decomposed by a smoothing function, and the complex nonlinear sequence is decomposed into two parts: an approximately integrable quantity and an integrable quantity, which is convenient for separate processing and improves the prediction accuracy. The approximate integrable quantity is subjected to trigonometric function periodic conversion processing, and the irregular time series is converted into a periodic function combination with clear mathematical characteristics, which enhances the predictability of the signal. The periodic feature sequence and the integrable quantity are subjected to linear merging transformation processing by weight coefficients, and the organic combination of periodic changes and random fluctuations is realized, so that the prediction model can capture the regularity and random characteristics of the signal at the same time. The combined feature sequence is trained and processed by a deep neural network, and the advantages of the convolutional neural network in extracting local features and the long short-term memory network in capturing long-term dependencies are fully utilized to construct a model with strong prediction ability. In the real-time prediction stage, the accuracy and practicality of the prediction results are ensured by preprocessing and anti-normalization. The overall solution significantly improves the prediction accuracy of frequency modulation instructions through the system design of data preprocessing, feature extraction, model training and prediction output. Experimental verification shows that the prediction RMSE of frequency modulation instructions 1 and 2 is 1.2 and 3.3 respectively, which is significantly improved compared with 9.9 and 8.7 of the traditional model, providing a reliable decision-making basis for the efficient operation of supercapacitor and lithium battery hybrid energy storage system. This improvement in prediction accuracy helps to optimize the power allocation strategy of the energy storage system, improve the frequency modulation performance of the system, extend the service life of the equipment, and reduce operating costs.

[0019] In a specific embodiment, the process of executing step S101 may specifically include the following steps: (1) Perform missing value detection on the original frequency modulation instruction sequence to obtain the missing value position information, and then perform linear interpolation on the missing value position information to obtain the completed sequence; (2) Perform outlier detection on the completed sequence to obtain an outlier labeled sequence, and perform median replacement on the outlier labeled sequence to obtain a denoised sequence; (3) Divide the denoised sequence into time windows to obtain time series segments of fixed length, and perform overlapping and sliding processing on the time series segments of fixed length to obtain sliding sequence samples; (4) Performing amplitude statistical processing on the sliding sequence samples to obtain the amplitude range parameter, and performing min-max normalization processing on the sliding sequence samples according to the amplitude range parameter to obtain a normalized sequence; (5) The normalized sequence is processed by Fourier transform to obtain frequency characteristic parameters, and the normalized sequence is subjected to bandpass filtering according to the frequency characteristic parameters to obtain a standardized frequency modulation sequence.

[0020] Specifically, missing value detection is performed, and the positions where the values ​​are empty, zero, or beyond the range of physical meaning are identified by scanning the data points in the original frequency modulation instruction sequence. The missing value detection adopts the sliding time window scanning method, and the window size is set to 10 seconds. The data in each time window is checked for continuity. When the data at a certain time point is found to be missing or abnormal, the index position of the time point is recorded in the missing value position information table. After obtaining the missing value position information, the linear interpolation method is used to repair the data. The specific method is to calculate the interpolation value according to the proportional relationship of the time interval based on the valid data points before and after the missing point. Linear interpolation ensures the continuity of the data and avoids mutations between the completed data and the original data. Outlier detection is performed on the completed sequence, and the box plot method is used to identify outliers. First, the quartiles of the sequence are calculated, including the first quartile Q1, the median Q2, and the third quartile Q3, and the interquartile range IQR = Q3 - Q1 is calculated. Data points that are less than (Q1 - 1.5×IQR) or greater than (Q3 + 1.5×IQR) are marked as outliers, and an outlier marker sequence is generated. For the marked outliers, the median replacement method is used to process them. The specific method is to calculate the median of all normal data points in the 15-second time window before and after the outlier point, and replace the outlier with the median. The median replacement method is insensitive to extreme values ​​and can effectively maintain the overall distribution characteristics of the data.

[0021] After the outlier processing is completed, the denoised sequence is divided into time windows. Considering the dynamic characteristics of the frequency modulation instruction, the time window length is set to 5 minutes. This length can capture the short-term fluctuation characteristics without losing important trend information. When dividing the window, the overlapping sliding method is adopted, and the overlap rate between adjacent windows is set to 75%, that is, the sliding step length is 75 seconds each time. Overlapping sliding can increase the correlation between samples and provide more training data. For each time window, the complete time series data segment in the window is extracted to form a fixed-length time series segment. Through overlapping sliding processing, these time series segments are organized into sliding sequence sample sets.

[0022] Perform amplitude statistical analysis on the sliding sequence samples and calculate the numerical distribution characteristics of the entire sample set. First, the maximum value Vmax and the minimum value Vmin in the sample set are counted to calculate the span of the numerical range. At the same time, the mean μ and the standard deviation σ are calculated, and these statistics constitute the amplitude range parameters. Based on these parameters, the sliding sequence samples are normalized by min-max, and all data are mapped to the interval [0,1]. The mathematical expression of normalization is: for the original value x, the normalized value is equal to (x-Vmin) / (Vmax-Vmin). Normalization eliminates the dimension effect and makes the data of different time periods comparable. Perform frequency domain analysis on the normalized sequence, and use the fast Fourier transform (FFT) method to convert the time domain signal to the frequency domain. The spectrum of the signal is obtained by FFT calculation, and the amplitude and phase information of each frequency component are analyzed to obtain the frequency characteristic parameters. The frequency characteristic parameters include the frequency value of the main frequency component, the corresponding amplitude size and phase angle. The bandpass filter is designed according to the frequency characteristic parameters. The passband range of the filter covers the main frequency components of the signal, while filtering out high-frequency noise and very low-frequency drift. The signal after filtering is the standardized frequency modulation sequence.

[0023] Taking the actual frequency modulation data of a thermal power plant as an example, the frequency modulation instruction data is collected at intervals of 1 second within a sampling period of one hour, and a total of 3600 data points are obtained. Through missing value detection, it is found that there are 15 missing values, located at time points such as 123 seconds and 256 seconds. Linear interpolation is used for these positions. For example, the missing value at 123 seconds is interpolated according to the value of 48.5MW at 122 seconds and the value of 49.2MW at 124 seconds, and the supplementary value at 123 seconds is 48.85MW. The outlier detection analysis shows that the Q1 of the original data is -35MW, Q3 is 35MW, and the IQR is 70MW. 25 outlier points are detected, such as -85MW at 789 seconds, which is beyond the normal range. It is replaced by the median -42MW in the 15-second window before and after this point. The processed sequence is divided into 5-minute windows to obtain 12 basic time windows, which are expanded to form 45 sample segments through sliding processing with a 75% overlap rate. Statistical analysis of these samples revealed a maximum value of 50MW, a minimum value of -50MW, a mean value close to 0MW, and a standard deviation of 28MW. The values ​​were mapped to the [0,1] interval through min-max normalization. FFT analysis found that the main frequency components were concentrated between 0.1Hz and 0.3Hz, and the passband range of the bandpass filter was designed to be 0.08Hz to 0.35Hz, filtering out slow drifts below 0.08Hz and high-frequency noise above 0.35Hz. The processed standardized FM sequence showed good periodicity and stability.

[0024] In a specific embodiment, the process of executing step S102 may specifically include the following steps: (1) The standardized frequency modulation sequence is processed by the Gaussian kernel function to obtain a smooth sequence, and the smooth sequence is processed by adaptively adjusting the window length to obtain the kernel function parameters; (2) The smoothed sequence is processed by convolution operation according to the kernel function parameters to obtain an approximate integrable quantity, and the standardized frequency modulation sequence and the approximate integrable quantity are subjected to difference calculation to obtain an integrable quantity.

[0025] Specifically, the Gaussian kernel function is used to smooth the standardized frequency modulation sequence. The basic form of the Gaussian kernel function is a bilateral exponential decay function, which is bell-shaped in the time domain. The reason for choosing the Gaussian kernel function as a smoothing tool is that it has good local properties and smoothing characteristics, and the processing of the signal will not introduce phase distortion. During the processing of the Gaussian kernel function, the weighted influence of other points in its neighborhood is considered for the processing of each data point. The weight coefficient decays in a Gaussian distribution as the distance from the center point increases, thereby obtaining a preliminary smoothed sequence. The window length is adaptively adjusted for the smoothed sequence in order to obtain the optimal kernel function parameters. The window length directly affects the degree of smoothing. A window that is too long will lead to over-smoothing and loss of effective fluctuation information, while a window that is too short cannot effectively suppress noise. The adaptive adjustment process is based on the local characteristics of the signal and uses a cross-validation method to determine the optimal parameters by comparing the smoothing effects under different window lengths. For areas with large fluctuations, the window length is automatically reduced to retain the rapidly changing features; for areas with gentle changes, the window length is appropriately increased to enhance the smoothing effect.

[0026] After determining the kernel function parameters, the smooth sequence is processed by convolution operation. The essence of convolution operation is to use Gaussian kernel function and signal for sliding weighted average, and the kernel function parameters determine the weighting method and range. Through the convolution operation, the high-frequency components of the signal are attenuated, while the main trend is retained to form an approximate integrable quantity. The approximate integrable quantity reflects the basic change trend of the frequency modulation instruction, has good mathematical properties, and is convenient for subsequent periodic processing. Finally, the difference between the standardized frequency modulation sequence and the approximate integrable quantity is calculated to obtain the integrable quantity. The integrable quantity contains the fast fluctuation components that are smoothed and filtered out in the original sequence. Although these components do not have good integral properties, they contain important local feature information. Through this decomposition method, the complex frequency modulation instruction sequence is divided into two processable parts, laying the foundation for subsequent feature extraction and predictive modeling.

[0027] Take the specific data processing as an example: the standardized frequency modulation sequence at a certain moment contains 100 data points within 10 seconds, and the value range is between [-1,1]. After initial processing with the Gaussian kernel function, the kernel function with a standard deviation of 0.5 seconds is selected for smoothing to obtain a preliminary smoothed sequence. In the adaptive adjustment process, the data is divided into 5 sub-segments, each with a 2-second period, and the local fluctuation intensity is calculated separately. For the sub-segment with a fluctuation intensity of 0.8, the kernel function standard deviation is adjusted to 0.3 seconds; for the sub-segment with a fluctuation intensity of 0.2, the kernel function standard deviation is adjusted to 0.7 seconds. The convolution operation with the adjusted parameters shows obvious trend characteristics, and the amplitude is about 80% of the original sequence. Subtracting the original sequence from the approximate integrable quantity, the amplitude of the non-integrable quantity is about 20% of the original sequence, which is mainly reflected in the high-frequency fluctuation component. The entire processing process maintains the conservation of signal energy and provides reliable input data for subsequent periodic conversion.

[0028] In a specific embodiment, the process of executing step S103 may specifically include the following steps: (1) Performing basic period estimation processing on the approximate integrable quantity to obtain period parameters, and performing segmentation processing on the approximate integrable quantity according to the period parameters to obtain segmented sequences; (2) Perform harmonic decomposition processing on the segmented sequence to obtain fundamental and harmonic components, and perform sine-cosine function fitting processing on the fundamental and harmonic components to obtain fitting coefficients; (3) Perform superposition calculation on the fitting coefficients to obtain the periodic characteristic sequence.

[0029] Specifically, the main periodic characteristics of the sequence are determined by the basic period estimation process, and the correlation coefficient of the sequence under different time delays is calculated by the autocorrelation analysis method. When the first significant peak of the correlation coefficient appears, the corresponding time delay is the basic period. According to the obtained period parameters, the approximate integrable quantity is divided into equal length segments, and the length of each segment is equal to the basic period, ensuring that the segmentation points are located at similar phases of the waveform, so as to obtain multiple complete segmented sequences of periods. The harmonic decomposition of the segmented sequence is to extract the contribution of different frequency components. The fundamental component corresponds to the basic period of the sequence, and its frequency is 1 / T (T is the basic period). The harmonic component is an integer multiple of the fundamental frequency, and the fifth-order harmonic is usually considered. Through the Fourier analysis method, each segmented sequence is decomposed into the fundamental wave and multiple harmonic components. Sine and cosine functions are fitted to these components respectively to solve the amplitude and phase parameters. The fitting process uses the least squares method to obtain the corresponding sine term and cosine term coefficients for each frequency component, and these coefficients together constitute the fitting coefficient set. Finally, the fitting results of all frequency components are superimposed to achieve the reconstruction of the periodic characteristic sequence. The importance of each component is considered during the superposition calculation. Usually, the weight of the fundamental component is the largest, and the weight of the higher harmonics decreases as the frequency increases. The reconstructed periodic characteristic sequence retains the main periodic change characteristics of the original sequence, providing a reliable basis for subsequent prediction analysis.

[0030] Take the actual frequency modulation data processing as an example: perform autocorrelation analysis on a certain segment of approximately integrable sequence, and by calculating the correlation coefficient of different time delays (0-600 seconds), find the first significant peak at 180 seconds, with a correlation coefficient of 0.85, and determine the basic period to be 180 seconds. According to this period parameter, divide the sequence into multiple segments, each containing 180 data points (sampling frequency 1Hz). Perform harmonic analysis on a single segmented sequence, and obtain the fundamental frequency of 0.0056Hz, and the 2nd-5th harmonic frequencies of 0.0112Hz, 0.0168Hz, 0.0224Hz, and 0.0280Hz, respectively. The amplitude of the fundamental component is 0.6, and the phase angle is 45 degrees. The amplitudes of the second to fifth harmonics are 0.3, 0.15, 0.08, and 0.04, respectively, and the phase angles are 30 degrees, 60 degrees, 90 degrees, and 120 degrees, respectively. These components are superimposed according to the obtained amplitude and phase to reconstruct a periodic characteristic sequence. The correlation coefficient between this sequence and the original data reaches 0.92, indicating that the reconstruction effect is good.

[0031] In a specific embodiment, the process of executing step S104 may specifically include the following steps: (1) Perform importance assessment on the periodic characteristic sequence to obtain the periodic characteristic weight, and perform volatility analysis on the integrable quantity to obtain the volatility characteristic weight; (2) The periodic characteristic sequence is weighted according to the periodic characteristic weight to obtain a weighted periodic sequence, and the integrable quantity is weighted according to the fluctuation characteristic weight to obtain a weighted integrable sequence; (3) Perform linear combination processing on the weighted periodic sequence and the weighted non-integrable sequence to obtain the combined characteristic sequence.

[0032] Specifically, weight allocation and combination of periodic feature sequences and non-integrable quantities are important links for accurate prediction. The weight of the periodic feature sequence is determined by importance evaluation, and the evaluation method is based on the degree of explanation of the periodic feature sequence for the changes in the original sequence. Specifically, the mutual correlation coefficient between the periodic feature sequence and the original standardized frequency modulation sequence is calculated, the energy distribution of the periodic feature sequence is analyzed, and its prediction ability on different time scales is evaluated. At the same time, the volatility analysis of the non-integrable quantity is performed, and its statistical characteristics such as root mean square value, kurtosis, skewness, etc. are calculated to evaluate the intensity and distribution characteristics of random fluctuations, thereby obtaining the volatility feature weight. According to the weight coefficients obtained by the evaluation, the periodic feature sequence and the non-integrable quantity are weighted respectively. For the periodic feature sequence, the periodic feature weight is applied to each data point of the sequence to obtain a weighted periodic sequence. This process highlights the part of the sequence with significant periodicity and enhances the expression of regular characteristics. For the non-integrable quantity, the volatility feature weight is applied to each volatility component in the sequence to obtain a weighted non-integrable sequence. This weighting process retains the meaningful random fluctuation information in the sequence while suppressing irrelevant noise interference.

[0033] Finally, the weighted periodic sequence and the weighted non-integrable sequence are linearly combined, and the normalized weight coefficient is used to ensure that the combined sequence maintains an appropriate numerical range. The linear combination process achieves the optimal fusion of periodic characteristics and random fluctuation characteristics through fine adjustment of weights. The obtained combined feature sequence not only retains the main change law of the frequency modulation instruction, but also contains the necessary fluctuation information.

[0034] Taking specific data processing as an example: the importance of the periodic feature sequence of a certain section of frequency modulation data is evaluated. By calculating the correlation coefficient with the original sequence, a correlation of 0.85 is obtained, and the energy share reaches 75%. Based on this, the weight of the periodic feature is determined to be 0.7. The volatility analysis of the non-integrable quantity shows that its root mean square value is 0.3, the kurtosis is 3.2, and the skewness is 0.1, indicating that the volatility distribution is relatively uniform, and the volatility feature weight is determined to be 0.3. A weight coefficient of 0.7 is applied to the periodic feature sequence for weighting, and the part of the original sequence with an amplitude of ±0.8 is adjusted to ±0.56. A weight coefficient of 0.3 is applied to the non-integrable quantity for weighting, and the part of the original fluctuation amplitude of ±0.4 is adjusted to ±0.12. Finally, the two weighted sequences are linearly combined, and the resulting combined feature sequence has moderate random fluctuations superimposed within the periodic variation range, and the overall variation range of the sequence is controlled between ±0.7, which not only maintains the main characteristics of the original signal, but also has good predictive value.

[0035] In a specific embodiment, the process of executing step S105 may specifically include the following steps: (1) Performing training set partitioning processing on the combined feature sequence to obtain a training sample set, and performing batch size setting processing on the training sample set to obtain a data batch; (2) The data batches are processed through a three-layer convolutional neural network to obtain a feature mapping sequence, and the feature mapping sequence is processed through a bidirectional long short-term memory network to obtain time series features; (3) The timing features are processed through a fully connected layer to obtain a frequency modulation instruction prediction model.

[0036] Specifically, in the frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage, deep neural network training of combined feature sequences is the core link of building a prediction model. First, the combined feature sequence is divided into training sets according to the time order, and the sliding window method is used to set the input window length to 30 minutes and the prediction window length to 5 minutes. The construction of the training sample set adopts a ratio division method of 70% for training, 15% for verification, and 15% for testing. The batch size is set for the training sample set. Considering the memory limitation and training efficiency, the batch size is set to 64, and the training samples are organized into corresponding data batches through the data loader. The constructed three-layer convolutional neural network is used to extract local features. The first layer of convolution uses 32 convolution kernels of size 3×1 and a step size of 1 to extract features from the input sequence; the second layer of convolution uses 64 convolution kernels of size 3×1 and a step size of 1 to increase the richness of features; the third layer of convolution uses 128 convolution kernels of size 3×1 and a step size of 1 to further extract deep features. Each convolution layer is followed by a batch normalization layer and a ReLU activation function, and a maximum pooling layer is used to reduce the data dimension. The feature map sequence output by the convolutional network contains the local temporal features of the frequency modulation instructions.

[0037] The feature mapping sequence is processed using a bidirectional long short-term memory network, which contains two layers of LSTM structure, each layer contains 256 hidden units. The bidirectional structure can simultaneously consider the forward and backward dependencies of the sequence and improve the ability to extract time series features. The LSTM unit contains an input gate, a forget gate, and an output gate, which can effectively handle long-term dependency problems. Through the processing of the bidirectional LSTM, a feature vector encoding the long-term and short-term time series relationship is obtained. Finally, a three-layer fully connected network is used for feature mapping. The first layer contains 512 neurons, the second layer contains 256 neurons, and the number of neurons in the third layer matches the prediction target dimension. A Dropout layer is added between the fully connected layers, and the loss rate is set to 0.3 to prevent overfitting. The last layer uses a linear activation function to output the prediction result.

[0038] Take the actual training process as an example: the frequency modulation data of a thermal power plant for one week contains 10080 sampling points (sampling interval is 1 minute), and 8640 training samples are obtained through sliding windows. After setting the batch size to 64, 135 training batches are formed. After the first layer of convolution processing, the feature map size is 28×32; after the second layer of convolution, it becomes 26×64; the third layer of convolution outputs a 24×128 feature mapping sequence. After bidirectional LSTM processing, a 512-dimensional time series feature vector is obtained. After mapping in the fully connected layer, a prediction sequence with a resolution of 5 minutes is finally output. Evaluated on the test set, the model predicts the frequency modulation instructions within the next 5 minutes with a root mean square error of 1.2MW.

[0039] In a specific embodiment, the process of executing step S106 may specifically include the following steps: (1) Preprocess the real-time frequency modulation instruction data to obtain a real-time standardized sequence, and perform forward calculation processing on the real-time standardized sequence through the frequency modulation instruction prediction model to obtain the original prediction value; (2) The original prediction value is subjected to constraint checking to obtain a valid prediction result, and the valid prediction result is subjected to inverse normalization transformation to obtain a frequency modulation instruction prediction value.

[0040] Specifically, the processing of real-time frequency modulation instruction data is the last step to achieve accurate prediction. First, the real-time frequency modulation instruction data needs to be preprocessed. This process is consistent with the data processing in the training phase, including data cleaning, normalization and feature extraction. The received real-time data is detected for missing values ​​and processed for outliers. The min-max normalization is performed using the same parameters as in the training phase to ensure the uniformity of the data format and value range, and obtain a real-time standardized sequence. The standardized sequence is input into the frequency modulation instruction prediction model for forward calculation. The model passes through the convolution layer, the loop layer and the fully connected layer in turn according to the feature mapping relationship established during training to calculate the original prediction value. Constraint checking of the original prediction value output by the model is an important step to ensure the validity of the prediction result. Constraint checking includes two aspects: physical constraints and logical constraints. Physical constraints ensure that the prediction value does not exceed the actual regulation capability of the system, such as checking whether the prediction value is within the maximum regulation rate limit of the thermal power unit; logical constraints ensure the continuity and smoothness of the prediction value to avoid unreasonable jumps. The prediction results that pass the constraint check are inversely normalized, and the values ​​in the normalized space are mapped back to the original value domain to obtain the final frequency modulation instruction prediction value.

[0041] Take the actual prediction process as an example: the real-time frequency regulation command data received at a certain moment is preprocessed, and the original value range is between -45MW and +52MW and normalized to the interval of 0-1. The original prediction value of 0.7 is obtained through the forward calculation of the prediction model. Through the physical constraint check, it is found that the corresponding change rate of this value is 2.5MW / minute, which does not exceed the regulation limit of 3MW / minute of the unit. At the same time, through the logical constraint check, it is confirmed that the change amplitude of the predicted value and the previous moment is within a reasonable range. Finally, through the inverse normalization transformation, 0.7 is mapped back to the original dimension to obtain a frequency regulation command prediction value of 35MW, which not only meets the physical constraints of the system, but also has good time series continuity.

[0042] The frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage in the embodiment of the present application is described above. The frequency modulation instruction prediction system for supercapacitor coupled lithium battery energy storage in the embodiment of the present application is described below. Figure 2 In the embodiment of the present application, an embodiment of a frequency modulation instruction prediction system for supercapacitor coupled lithium battery energy storage includes: The processing module 201 is used to perform data cleaning and normalization processing on the original frequency modulation instruction sequence to obtain a standardized frequency modulation sequence; A decomposition module 202, configured to decompose the standardized frequency modulation sequence through a smoothing function to obtain an approximate integrable quantity and an integrable quantity; A conversion module 203 is used to convert the approximate integrable quantity into a periodic conversion process through a trigonometric function to obtain a periodic characteristic sequence, wherein the periodic characteristic sequence is composed of a plurality of sine and cosine functions of different frequencies; A transformation module 204 is used to perform a linear merging transformation process on the periodic feature sequence and the incommensurable quantity through a weight coefficient to obtain a combined feature sequence; A training module 205 is used to train the combined feature sequence through a deep neural network to obtain a frequency modulation instruction prediction model, wherein the frequency modulation instruction prediction model includes a three-layer convolutional neural network, a bidirectional long short-term memory network and a fully connected layer; The prediction module 206 is used to perform prediction processing on the real-time frequency modulation instruction data through the frequency modulation instruction prediction model to obtain a prediction result, and perform denormalization processing on the prediction result to obtain a frequency modulation instruction prediction value.

[0043] Through the cooperation of the above components, the missing values, outliers and noise interference in the data are effectively removed by cleaning and normalizing the original frequency modulation instruction sequence, ensuring the reliability of data quality. The standardized frequency modulation sequence is decomposed by smoothing function, and the complex nonlinear sequence is decomposed into two parts: approximate integrable quantity and incomprehensible quantity, which is convenient for separate processing and improves the prediction accuracy. The approximate integrable quantity is transformed into a periodic function combination with clear mathematical characteristics by trigonometric function periodic conversion, which enhances the predictability of the signal. The periodic feature sequence and the incomprehensible quantity are transformed into a linear combination of periodic changes and random fluctuations by weight coefficients, which realizes the organic combination of periodic changes and random fluctuations, so that the prediction model can capture the regularity and random characteristics of the signal at the same time. The combined feature sequence is trained by deep neural network, which makes full use of the advantages of convolutional neural network in extracting local features and long short-term memory network in capturing long-term dependencies, and constructs a model with strong prediction ability. In the real-time prediction stage, the accuracy and practicality of the prediction results are ensured by preprocessing and anti-normalization. The overall solution significantly improves the prediction accuracy of frequency modulation instructions through the system design of data preprocessing, feature extraction, model training and prediction output. Experimental verification shows that the prediction RMSE of frequency modulation instructions 1 and 2 is 1.2 and 3.3 respectively, which is significantly improved compared with 9.9 and 8.7 of the traditional model, providing a reliable decision-making basis for the efficient operation of supercapacitor and lithium battery hybrid energy storage system. This improvement in prediction accuracy helps to optimize the power allocation strategy of the energy storage system, improve the frequency modulation performance of the system, extend the service life of the equipment, and reduce operating costs.

[0044] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage, characterized in that: The frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage includes: Perform data cleaning and normalization on the original frequency modulation instruction sequence to obtain a standardized frequency modulation sequence; Decomposing the standardized frequency modulation sequence by a smoothing function to obtain an approximate integrable quantity and an integrable quantity; The approximate integrable quantity is subjected to periodic conversion processing by trigonometric functions to obtain a periodic characteristic sequence, wherein the periodic characteristic sequence is composed of a plurality of sine and cosine functions of different frequencies; Performing linear merging transformation processing on the periodic characteristic sequence and the non-integrable quantity through weight coefficients to obtain a combined characteristic sequence; The combined feature sequence is trained through a deep neural network to obtain a frequency modulation instruction prediction model, wherein the frequency modulation instruction prediction model includes a three-layer convolutional neural network, a bidirectional long short-term memory network and a fully connected layer; The real-time frequency modulation instruction data is predicted by the frequency modulation instruction prediction model to obtain a prediction result, and the prediction result is denormalized to obtain a frequency modulation instruction prediction value.

2. The frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage according to claim 1, characterized in that: The data cleaning and normalization processing of the original frequency modulation instruction sequence to obtain a standardized frequency modulation sequence includes: Performing missing value detection processing on the original frequency modulation instruction sequence to obtain missing value position information, and performing linear interpolation processing on the missing value position information to obtain a completed sequence; Performing outlier detection processing on the completed sequence to obtain an outlier marker sequence, and performing median replacement processing on the outlier marker sequence to obtain a denoised sequence; Performing time window division processing on the denoised sequence to obtain time sequence segments of fixed length, and performing overlapping sliding processing on the time sequence segments of fixed length to obtain sliding sequence samples; Performing amplitude statistical processing on the sliding sequence samples to obtain an amplitude range parameter, and performing min-max normalization processing on the sliding sequence samples according to the amplitude range parameter to obtain a normalized sequence; The normalized sequence is processed by Fourier transform to obtain frequency characteristic parameters, and the normalized sequence is subjected to bandpass filtering according to the frequency characteristic parameters to obtain the standardized frequency modulation sequence.

3. The frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage according to claim 1, characterized in that: Decomposing the standardized frequency modulation sequence by a smoothing function to obtain approximate integrable quantities and integrable quantities includes: Processing the standardized frequency modulation sequence by a Gaussian kernel function to obtain a smoothed sequence, and processing the smoothed sequence by adaptively adjusting the window length to obtain kernel function parameters; The smoothed sequence is subjected to convolution operation according to the kernel function parameters to obtain an approximate integrable quantity, and the standardized frequency modulation sequence and the approximate integrable quantity are subjected to difference calculation to obtain an integrable quantity.

4. The frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage according to claim 1, characterized in that: The approximate integrable quantity is subjected to periodic conversion processing by a trigonometric function to obtain a periodic characteristic sequence, wherein the periodic characteristic sequence is composed of a plurality of sine and cosine functions of different frequencies, including: Performing basic period estimation processing on the approximate integrable quantity to obtain period parameters, and performing segmentation processing on the approximate integrable quantity according to the period parameters to obtain a segmented sequence; Performing harmonic decomposition processing on the segmented sequence to obtain fundamental wave and harmonic wave components, and performing sine-cosine function fitting processing on the fundamental wave and harmonic wave components to obtain fitting coefficients; The fitting coefficients are subjected to superposition calculation processing to obtain the periodic characteristic sequence.

5. The frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage according to claim 1, characterized in that: The linear merging and transformation processing of the periodic feature sequence and the incommensurable quantity by weight coefficients to obtain a combined feature sequence includes: Performing importance assessment processing on the periodic characteristic sequence to obtain a periodic characteristic weight, and performing volatility analysis processing on the non-integrable quantity to obtain a volatility characteristic weight; Performing weighted processing on the periodic characteristic sequence according to the periodic characteristic weight to obtain a weighted periodic sequence, and performing weighted processing on the integrable quantity according to the fluctuation characteristic weight to obtain a weighted integrable sequence; The weighted periodic sequence and the weighted non-integrable sequence are linearly combined to obtain the combined characteristic sequence.

6. The frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage according to claim 1, characterized in that: The combined feature sequence is trained through a deep neural network to obtain a frequency modulation instruction prediction model, wherein the frequency modulation instruction prediction model includes a three-layer convolutional neural network, a bidirectional long short-term memory network and a fully connected layer, including: Performing training set division processing on the combined feature sequence to obtain a training sample set, and performing batch size setting processing on the training sample set to obtain a data batch; Processing the data batches through a three-layer convolutional neural network to obtain a feature mapping sequence, and processing the feature mapping sequence through a bidirectional long short-term memory network to obtain a time series feature; The timing features are processed through a fully connected layer to obtain the frequency modulation instruction prediction model.

7. The frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage according to claim 1, characterized in that: The real-time frequency modulation instruction data is predicted by the frequency modulation instruction prediction model to obtain a prediction result, and the prediction result is denormalized to obtain a frequency modulation instruction prediction value, including: Performing data preprocessing on the real-time frequency modulation instruction data to obtain a real-time standardized sequence, and performing forward calculation processing on the real-time standardized sequence through the frequency modulation instruction prediction model to obtain an original prediction value; The original prediction value is subjected to constraint checking to obtain a valid prediction result, and the valid prediction result is subjected to inverse normalization transformation to obtain the frequency modulation instruction prediction value.

8. A frequency modulation instruction prediction system for supercapacitor coupled lithium battery energy storage, used to implement the frequency modulation instruction prediction method for supercapacitor coupled lithium battery energy storage as described in any one of claims 1-7, characterized in that: The frequency modulation instruction prediction system for supercapacitor coupled lithium battery energy storage includes: A processing module is used to perform data cleaning and normalization processing on the original frequency modulation instruction sequence to obtain a standardized frequency modulation sequence; A decomposition module, used for decomposing the standardized frequency modulation sequence through a smoothing function to obtain an approximate integrable quantity and an integrable quantity; A conversion module, used for performing periodic conversion processing on the approximate integrable quantity through trigonometric functions to obtain a periodic characteristic sequence, wherein the periodic characteristic sequence is composed of a plurality of sine and cosine functions of different frequencies; A transformation module, used for performing a linear merging transformation process on the periodic feature sequence and the non-integrable quantity through a weight coefficient to obtain a combined feature sequence; A training module, used for training the combined feature sequence through a deep neural network to obtain a frequency modulation instruction prediction model, wherein the frequency modulation instruction prediction model includes a three-layer convolutional neural network, a bidirectional long short-term memory network and a fully connected layer; The prediction module is used to perform prediction processing on the real-time frequency modulation instruction data through the frequency modulation instruction prediction model to obtain a prediction result, and perform denormalization processing on the prediction result to obtain a frequency modulation instruction prediction value.

Images