A short-term load combined prediction method based on adaptive chirp modal decomposition and SSA-BiLSTM

Abstract

The application discloses a short-term load combined prediction method based on adaptive chirp mode decomposition and SSA-BiLSTM, which comprises the following steps: S1, setting the sparrow population number of SSA, the proportion of sparrows with different task division and the maximum iteration number; S2, initializing the to-be-optimized parameters in the solution space; S3, training the sample by constructing a BiLSTM model corresponding to the parameter value, and calculating the fitness value; S4, updating the sparrow population position according to the fitness value output by the BiLSTM, obtaining a new solution, and transmitting the corresponding solution, i.e., the parameter value, to the BiLSTM for training; S5, repeating the step S2 and the step S3 until the maximum iteration number is reached and the iteration is completed; and S6, outputting the optimal parameter value combination and establishing an optimal BiLSTM prediction model. The short-term load prediction precision is effectively improved by solving the technical problems, such as slow convergence, complicated parameter adjustment, poor stability and low prediction precision, of the load prediction method in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to load forecasting technology, and in particular to a short-term load combined forecasting method based on adaptive chirplet mode decomposition and SSA-BiLSTM. BACKGROUND

[0002] Currently, the methods related to load forecasting can be divided into the following three categories: first, the traditional forecasting methods represented by linear regression, ARIMA and exponential smoothing. Although the traditional method has a simple structure and is easy to implement, it performs poorly in nonlinear time series forecasting problems. Second, the artificial intelligence methods represented by machine learning and deep learning. This kind of method improves the processing ability of nonlinear problems. In recent years, with the improvement of computer computing power, especially the deep learning model represented by neural network, is widely used in load forecasting. Third, the combined forecasting method. The common idea of combined forecasting method is divided into two kinds: one is to combine heuristic optimization algorithm with single forecasting model to select and set the value of model hyperparameters. The literature uses particle swarm optimization algorithm to optimize deep belief network model. The literature proposes a prediction model based on genetic algorithm optimization BP neural network. Compared with single model, the prediction accuracy is improved. The second idea is to combine the signal decomposition technology in information field to preprocess the time series data to reduce the complexity of the data, and then make prediction. The literature proposes a short-term load forecasting model based on empirical mode decomposition (EMD) and LSTM combination. Compared with single LSTM model, the prediction accuracy is higher. The literature combines variational mode decomposition (VMD) and LSTM to reduce the prediction error.

[0003] For example, a combination type power load forecasting method disclosed in Chinese patent document CN103679287A has the technical problems of slow convergence speed, complicated and frequent parameter adjustment, low overall stability of the prediction system, etc., and cannot effectively improve the accuracy of short-term load forecasting. SUMMARY

[0004] The present application is to solve the technical problems of slow convergence, complicated parameter adjustment, poor stability and low prediction accuracy of the existing load forecasting method, effectively improve the accuracy of short-term load forecasting, and provide a short-term load combined forecasting method based on adaptive chirplet mode decomposition (ACMD) and sparrow search algorithm (SSA) optimization bidirectional long short-term memory network (BiLSTM).

[0005] In order to achieve the above purpose, the present application adopts the following technical methods:

[0006] A short-term load combined prediction method based on adaptive chirp mode decomposition and SSA-BiLSTM, characterized in that the scheme comprises the following steps:

[0007] Step S1: combining the load data collected by the power distribution network power consumption data collection system and the meteorological data collected by the meteorological station near the power distribution area to form an original data set, and dividing the original data set into 80% training data and 20% test data;

[0008] Step S2: for the load data in the training data, the historical load sequence data is decomposed by using the adaptive chirp mode decomposition (ACMD) method, and q sub-sequences are obtained; combined with the meteorological data characteristics in the test data set;

[0009] Step S3: for the q sub-sequences after decomposition, a bidirectional long short-term memory network (BiLSTM) prediction model is constructed respectively; the sparrow search algorithm (SSA) is used to optimize the q BiLSTM models constructed, the hyperparameter values of the optimized model are optimized, and the optimal prediction model is constructed according to the optimized values;

[0010] Step S4: according to the established optimal prediction model, each sub-sequence is predicted, the prediction results are superimposed and combined to obtain the final prediction results, and compared with other prediction methods, the prediction results are evaluated through relevant evaluation indexes, and the prediction effect of the model is analyzed.

[0011] Compared with other prediction methods, the prediction results are evaluated through relevant evaluation indexes, and the prediction effect of the model is analyzed. The sequence data generated by the original data after ACDM decomposition is input into the BiLSTM prediction model, the SSA algorithm is used to optimize the parameters of the prediction model, and finally the optimized model outputs the prediction results. The output results of the multiple optimized prediction models are superimposed and compared with the original data, and the prediction result accuracy can be evaluated.

[0012] As a preferred, the SSA-BiLSTM prediction model establishment process decomposes the original unbalanced sequence by using the ACDM algorithm, adopts the recursive algorithm framework for matching tracking, and processes the multi-mode signal with strong time-varying characteristics; the ACDM is used to process the historical power signal required for load prediction, and the specific formula is as follows:

[0013]

[0014] Where Q is the number of signal modes, Aq(t), fq(t) and respectively represent the instantaneous amplitude, instantaneous frequency and initial phase of the qth signal mode.

[0015] The power load is influenced by production and life, meteorological factors, economic factors and the like, presents certain volatility, non-stationary characteristics, but since the power consumption habit of the power user has certain regularity, therefore, the load also has obvious periodic fluctuation trend, the algorithm of the present application can better capture the periodic regularity of the load, will help to effectively improve the precision of the prediction.

[0016] Further, the above historical power signal processing formula is rewritten through modulation and demodulation, and is specifically as follows:

[0017]

[0018]

[0019] Wherein, bq(t) and dq(t) are demodulation operators, represents demodulation frequency (DF), is instantaneous amplitude; it can be found that when the demodulation frequency is equal to the instantaneous frequency, that is, when , the frequency modulation term in the demodulation signal is 0, and the pure amplitude modulation signal with the narrowest bandwidth is obtained.

[0020] Further, by minimizing the bandwidth of the demodulation signal, the ACMD algorithm can estimate the intermediate frequency, decompose the signal, and estimate the components of the signal one by one, and the optimization problem is solved by alternately updating the demodulation signal and the frequency function to realize the decomposition of the original signal one by one. After obtaining the first signal component, it is subtracted from the original signal, and then the remaining component is taken as the new initial signal to continue to decompose to obtain the second signal component. The update is repeated until all signal components are obtained, and is specifically as follows:

[0021]

[0022] Wherein, ||...||2 is L2 norm, used to represent distance; is the second derivative, and its physical meaning is the smoothness of the single mode curve after decomposition; λ is a penalty factor; and represent bandwidth. For the discrete signal of the time sequence t=t0,…,tN-1, the number of sampling points is N, and the discrete expression form of the above formula is:

[0023]

[0024] Wherein, s=[s(t0),…,s(t N-1 ) T , u q =[(b q ) T ,(d q ) T ] T , bq =[b q (t0),…,b q (t N-1 )] T ,d q =[d q (t0),…,d q (t N-1 )] T Λ = diag[Ω,Ω] is a block diagonal matrix, and Ω is a second-order difference matrix. This optimization problem is solved by alternately updating the demodulated signal and the frequency function to achieve the decomposition of the original signal one by one. After obtaining the first signal component, it is subtracted from the original signal, and then the remaining component is used as the new initial signal to continue decomposing to obtain the second signal component. This cycle is repeated until all signal components are obtained, reducing the frequency of parameter changes and improving the accuracy of the prediction system.

[0025] Furthermore, BiLSTM is composed of two LSTM layers, and the two LSTM layers are processed in both forward and backward directions. The forward layer is used to capture the feature sequence before the current time point, and the backward layer is used to capture the feature sequence from the future time point to the current time point.

[0026] Furthermore, the prediction model building process continuously updates the positions of various sparrows in the D-dimensional solution space and compares the fitness values ​​to seek a better position, i.e., to find the optimal solution of the objective function. The various sparrows include discoverers, joiners, and alarmers.

[0027] SSA is inspired by the foraging behavior of sparrows. Compared with particle swarm optimization and genetic algorithms, it has the advantages of high speed, high accuracy and fewer parameters, and has great potential for engineering applications.

[0028] Furthermore, in the above process, the discoverer's location is updated as follows:

[0029]

[0030] in, It represents the d-th dimension position of individual i in the t-th iteration of the population; a is a random number between 0 and 1; iter max is the maximum number of iterations for the population; Q is a random number that follows a Gaussian distribution; R2 is also a random number between 0 and 1; ST is the alarm threshold between 0.5 and 1; when the value is greater than the alarm value, the position will be updated randomly; when the value is less than the alarm value, the position will converge at the current position.

[0031] The formula for updating the position of new members is as follows:

[0032]

[0033] Among them, is the position of the current worst sparrow individual, is the position of the best individual; rand{-1,1} represents randomly taking -1 and 1; n is the total number of individuals. When i > n / 2, it means that the low-fitness joiner does not obtain food at the current position and needs to update the position; when i < n / 2, it means that it is near the optimal position and continues to forage near the current position.

[0034] The calculation formula for updating the position of the alarm caller is as follows:

[0035]

[0036] Among them, β is a random number obeying the Gaussian distribution; f i is the fitness value of the current individual, f b is the fitness value of the optimal individual in the current population, f w is the fitness value of the worst individual; K is a random number taking values between -1 and 1; ε is a very small non-zero number; if the alarm caller is at the current optimal position, it will flee towards the vicinity of the optimal position, and if it is not at the current optimal position, it will flee to the vicinity of the optimal position.

[0037] There are some hyperparameters in the BiLSTM algorithm. When using BiLSTM to predict power load, different values of hyperparameters will have a great impact on the prediction accuracy of the model. Manually adjusting parameters according to experience is time-consuming and laborious, and has strong uncertainty. Therefore, in order to improve the prediction accuracy, the powerful optimization ability of SSA is used to quickly and accurately determine the values of each hyperparameter of BiLSTM. In this paper, three hyperparameters that have a greater impact on BiLSTM are selected as the optimization objects, namely: learning rate (learning_rate), the number of units in the first hidden layer (numHiddenUnits1), and the number of units in the second hidden layer (numHiddenUnits2). The mean square error function (MSE) is used as the loss function of BiLSTM and also as the fitness function of SSA; the goal of SSA is to find a set of hyperparameter combinations that minimize the MSE of the BiLSTM training samples.

[0038] MSE represents the degree of difference between the predicted power and the actual power, and the formula is as follows:

[0039]

[0040] In the formula, h is the total number of samples in the training set, y i is the true value of the i-th sampling point, y i is the predicted value.

[0041] Preferably, the establishment process of the prediction model includes:

[0042] Step SA1: Set the number of sparrows in the SSA population, the proportion of sparrows in different task assignments, and the maximum number of iterations;

[0043] Step SA2: Initialize the parameters to be optimized in the solution space;

[0044] Step SA3: Use the corresponding parameter values ​​to construct a BiLSTM model to train the samples and calculate the fitness value;

[0045] Step SA4: Update the sparrow population position based on the fitness value transmitted from BiLSTM to obtain a new solution. Transmit the corresponding solution, i.e. the parameter values, to BiLSTM for training.

[0046] Step SA5: Repeat steps S2 and S3 until the iteration is complete;

[0047] Step SA6: Output the optimal combination of parameter values ​​and establish the optimal BiLSTM prediction model.

[0048] Therefore, the beneficial effects of the present invention are as follows:

[0049] Compared to the traditional EMD method, the ACMD method is better at decomposing nonlinear signals. It can decompose the original load sequence into a series of more stable and regular subsequences, and the number of subsequences after decomposition is smaller, which reduces the workload of prediction and improves the accuracy of prediction.

[0050] In the BiLSTM prediction process, introducing SSA to optimize the hyperparameters of BiLSTM can overcome the disadvantages of high uncertainty and poor performance caused by manually adjusting parameters.

[0051] The ACMD-SSA-BiLSTM combined prediction model predicts and optimizes simple subsequences derived from the original sequence, then combines the prediction results to obtain the final result. Compared with single models and their combined models, it offers better prediction performance and higher accuracy, demonstrating certain application advantages. Attached Figure Description

[0052] Figure 1 This invention provides a short-term load combination forecasting model based on ACMD-SSA-BiLSTM.

[0053] Figure 2 This is a diagram of the BiLSTM model structure used in this invention;

[0054] Figure 3 This is a flowchart of the optimization process for the SSA-BiLSTM prediction model in this invention;

[0055] Figure 4This is a diagram showing the results of the EMD algorithm decomposing the original load data in this invention;

[0056] Figure 5 This is a diagram showing the results of decomposing the original load data using the ACMD algorithm in this invention.

[0057] Figure 6 This is a distribution diagram of prediction errors under different decomposition methods in this invention;

[0058] Figure 7 This is a comparison chart of the model performance before and after SSA optimization in this invention;

[0059] Figure 8 This is a comparison chart of the prediction curves of six different prediction models in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0061] like Figure 1 As shown, the basic idea of ​​this invention is to combine the power distribution network's "electricity consumption data acquisition system" with wind speed, temperature, radiation intensity, relative humidity, and cloud cover collected by meteorological stations near the distribution area to form an original dataset. This original dataset is then divided into 80% training data and 20% test data. For the load data in the test dataset, the historical load sequence data is decomposed using the Adaptive Chirped Mode Decomposition (ACMD) method to obtain q sub-sequences. Combining the meteorological data characteristics in the test dataset, a Bidirectional Long Short-Term Memory (BiLSTM) prediction model is constructed for each of the q sub-sequences. The Sparrow Search Algorithm (SSA) is used to optimize the constructed q BiLSTM models, optimizing the hyperparameter values. Based on the optimized values, the optimal prediction model is constructed. The prediction results of each sub-sequence are superimposed and combined to obtain the final prediction result. The prediction results are compared with other prediction methods, and the prediction effect of the model is evaluated using relevant evaluation indicators. Specifically, the following steps are included:

[0062] A short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM is characterized by the following steps:

[0063] Step S1: Combine the load data collected by the power distribution network data acquisition system and the meteorological data collected by the meteorological station near the power distribution area to form the original dataset, and divide the original dataset into 80% training data and 20% test data.

[0064] Step S2: For the load data in the training data, the historical load sequence data is decomposed using the Adaptive Chirped Mode Decomposition (ACMD) method to obtain q subsequences; combined with the meteorological data features in the test dataset;

[0065] Step S3: Construct a bidirectional long short-term memory (BiLSTM) prediction model for each of the q decomposed subsequences; optimize the q BiLSTM models using the Sparrow Search Algorithm (SSA), optimize the hyperparameter values ​​of the models, and construct the optimal prediction model based on the optimized values;

[0066] Step S4: Predict each subsequence based on the established optimal prediction model, combine the prediction results to obtain the final prediction result, evaluate the prediction result through relevant evaluation indicators, and analyze the prediction effect of the model.

[0067] By collecting real-time data from the power distribution network data acquisition system, digital signals and power information are extracted from the data source. In practical applications, because the acquisition equipment is differentiated, real-time digital signals and power signals for each power zone can be obtained, with the received data measured in milliseconds. This method dynamically decomposes the load sequence data in real time based on the amount of data received per unit time, obtaining multiple sub-sequences. Through dynamic analysis of each sub-sequence, meteorological data characteristics are tested and obtained. Therefore, this method adopts adaptive chirped mode decomposition and SSA-BiLSTM, which can be directly used to obtain significantly improved recognition results under the premise of weak environmental data changes. The environmental mathematical signals and power load signals are analyzed to obtain the text information corresponding to the power load signals, and a target dataset is generated based on the text information.

[0068] like Figure 2 As shown, BiLSTM is composed of two LSTM layers, which consider the sequences in both the forward and backward directions. The forward layer is used to capture the feature sequence before the current time point, and the backward layer is used to capture the feature sequence from the future time point to the current time point.

[0069] Electricity load is influenced by factors such as production and daily life, meteorological conditions, and economic factors, exhibiting characteristics of fluctuation and non-stationarity. However, due to the regularity of electricity consumption habits among users, the load also shows a clear cyclical fluctuation trend. If the algorithm can effectively capture the cyclical pattern of the load, it will help improve the accuracy of prediction. This paper uses the ACMD algorithm to decompose the original non-stationary sequence. ACMD is an adaptive signal decomposition algorithm that uses a recursive frame matching pursuit, which can effectively handle multi-mode signals with strong time-varying characteristics.

[0070] The historical power signal required for load forecasting using ACMD can be expressed by the following formula:

[0071]

[0072] Where Q is the number of signal modes, Aq(t), fq(t), and Let represent the instantaneous amplitude, instantaneous frequency, and initial phase of the q-th signal mode, respectively. Through modulation and demodulation, the above equation can be rewritten as:

[0073]

[0074]

[0075] In the formula, bq(t) and dq(t) are demodulation operators. Indicates the demodulation frequency (DF). This is the instantaneous amplitude; it can be observed that when the demodulation frequency equals the instantaneous frequency, that is, when... At this point, the frequency modulation term in the demodulated signal is 0, resulting in a pure amplitude modulation signal with the narrowest bandwidth. By minimizing the bandwidth of the demodulated signal, the ACMD algorithm can estimate the intermediate frequency, decompose the signal, and estimate the signal components one by one. That is, for the q-th signal component, solve the following optimization problem:

[0076]

[0077] Where ||...||2 is the L2 norm, used to represent the distance; ″ is the second derivative, whose physical meaning is the smoothness of the single-mode curve after decomposition; λ is the penalty factor; and This represents the bandwidth. For a discrete signal with a time series t = t0, ..., tN-1 and N sampling points, the discrete expression of the above formula is:

[0078]

[0079] In the formula, s = [s(t0), ..., s(t)] N-1 )] T u q =[(b q ) T ,(d q ) T ] T b q =[b q (t0),…,b q (t N-1 )] T ,d q =[d q (t0),…,d q (t N-1 )] T Λ=diag[Ω,Ω] is a block diagonal matrix, and Ω is a second-order difference matrix.

[0080] This optimization problem is solved by alternately updating the demodulated signal and the frequency function to decompose the original signal one by one. After obtaining the first signal component, it is subtracted from the original signal, and then the remaining component is used as the new initial signal to continue decomposing to obtain the second signal component. This process is repeated until all signal components are obtained.

[0081] like Figure 4 , Figure 5 As shown, to compare the decomposition effects of ACMD and EMD, 1500 sample points were selected from the training data, and EMD and ACMD were used for decomposition respectively. In terms of the number of subsequences after decomposition, EMD produced 9 subsequences, while ACMD produced only 6, effectively reducing the computational burden of building prediction models for each subsequence separately. Comparing the high-frequency components of the two methods, it can be found that the high-frequency components of EMD are more chaotic and lack regularity, while the high-frequency components of ACMD are more regular and exhibit a certain periodicity. Comparing the low-frequency components, it can be found that both methods have large amplitudes and gentle trends in their low-frequency components, but the low-frequency trend of ACMD is closer to the original sequence. In summary, the subsequences after ACMD decomposition are more conducive to subsequent prediction modeling. A BiLSTM prediction model is used for the load prediction model, and the SSA algorithm is used to optimize the model.

[0082] SSA was inspired by the foraging behavior of sparrows.

[21] Compared to particle swarm optimization and genetic algorithms, the sparrow algorithm boasts advantages such as high speed, high accuracy, and fewer parameters, demonstrating strong potential for engineering applications. Sparrow flocks exhibit clear division of labor during foraging, with discoverers locating high-quality food and followers joining in. Unlike the traditional discoverer-follower model, the sparrow algorithm incorporates a reconnaissance and early warning mechanism. A certain proportion of sparrows in the population are selected as alarm listeners to detect attacks and food-stealing behavior; upon detecting danger, they abandon their current food source. By continuously updating the positions of various sparrow types within a D-dimensional solution space and comparing fitness values, a better position is sought, thus finding the optimal solution to the objective function.

[0083] The location of the discoverer has been updated as follows:

[0084]

[0085] In the formula, It represents the d-th dimension position of individual i in the t-th iteration of the population; a is a random number between 0 and 1; iter max is the maximum number of iterations for the population; Q is a random number following a Gaussian distribution; R² is also a random number between 0 and 1; ST is the alarm threshold between 0.5 and 1. When the value exceeds the alarm threshold, the position is updated randomly; when the value is less than the alarm threshold, the population converges at the current position.

[0086] The description of the updation of the joiner's position is as follows:

[0087]

[0088] where is the position of the current worst sparrow individual, is the position of the best individual; rand{-1,1} means randomly taking -1 and 1; n is the total number of individuals. When i > n / 2, it means that the joiner with low fitness has not obtained food at the current position and needs to update the position. When i < n / 2, it means that it is near the optimal position and continues to forage near the current position.

[0089] A certain proportion of individuals in each generation of the population are selected as alarmers. The description of the updation of the alarmers is:

[0090]

[0091] where β is a random number obeying the Gaussian distribution; f i is the fitness value of the current individual, f b is the fitness value of the optimal individual in the current population, f w is the fitness value of the worst individual; K is a random number taking values between -1 and 1; ε is a very small non-zero number. It is easy to find that if the alarmer is at the current optimal position, it will flee towards the vicinity of the optimal position. If it is not at the current optimal position, it will flee to the vicinity of the optimal position.

[0092] There are some hyperparameters in the BiLSTM algorithm. When using BiLSTM to predict power load, different values of hyperparameters will have a great impact on the prediction accuracy of the model. Manually adjusting parameters according to experience is time-consuming and laborious, and has strong uncertainty. Therefore, in order to improve the prediction accuracy, the powerful optimization ability of SSA is used to quickly and accurately determine the values of each hyperparameter of BiLSTM. In this paper, three hyperparameters that have a greater impact on BiLSTM are selected as the optimization objects, namely: learning rate (learning_rate), the number of units in the first hidden layer (numHiddenUnits1), and the number of units in the second hidden layer (numHiddenUnits2). The mean square error function (MSE) is used as the loss function of BiLSTM and also as the fitness function of SSA. The goal of SSA is to find a set of hyperparameter combinations that minimize the MSE of the BiLSTM training samples.

[0093] MSE represents the degree of difference between the predicted power and the actual power. The formula is as follows:

[0094]

[0095] where h is the total number of samples in the training set, y iLet y be the true value of the i-th sampling point. i These are predicted values.

[0096] like Figure 6 As shown, by comparing the prediction error distributions of three prediction methods—BiLSTM, EMD-BiLSTM, and ACMD-BiLSTM—it was found that the prediction error of ACMD-BiLSTM was 40% of that of BiLSTM, and the prediction error of EMD-BiLSTM was 60%. Figure 7 As shown, the single BiLSTM model predicts a large number of samples with significant discrepancies between the predicted and actual values, while the SSA-BiLSTM stacked model shows almost no prediction bias compared to the actual values; for example... Figure 8 As shown, this scheme performs better in decomposing nonlinear signals, decomposing the original load sequence into a series of more stable and regular subsequences with fewer subsequences, reducing the workload of prediction while improving the accuracy of prediction. By superimposing the ACMD-SSA-BiLSTM combined prediction model, the simple subsequences decomposed from the original sequence are predicted and optimized separately, and the prediction results are superimposed to obtain the final result. Compared with the single model and its combination model, the prediction effect is better and the accuracy is higher.

[0097] like Figure 3 As shown, the process of establishing the SSA-BiLSTM optimization prediction model includes:

[0098] Step SA1: Set the number of sparrows in the SSA population, the proportion of sparrows with different task assignments, and the maximum number of iterations. Based on the target location information, determine the type of sparrows with different tasks.

[0099] Step SA2: Initialize the parameters to be optimized in the solution space;

[0100] Step SA3: Use the corresponding parameter values ​​to construct a BiLSTM model to train the samples and calculate the fitness value;

[0101] Step SA4: Update the sparrow population position based on the fitness value transmitted from BiLSTM to obtain a new solution. Transmit the corresponding solution, i.e. the parameter values, to BiLSTM for training.

[0102] Step SA5: Repeat steps S2 and S3 until the maximum number of iterations is reached and the iteration is complete;

[0103] Step SA6: Output the optimal combination of parameter values ​​and establish the optimal BiLSTM prediction model.

[0104] Using the above strategy, all prediction results are highly accurate for specific division of labor and data types. For power load affected by external forces such as production, life, weather, and economic factors, this prediction model captures the periodic patterns of power load fluctuations and imbalances, effectively handling multimode signals with strong time-varying characteristics.

[0105] The above description of the structure, features, and effects of the present invention is based on the embodiments shown in the figures. However, the above are only preferred embodiments of the present invention. It should be noted that the technical features involved in the above embodiments and their preferred methods can be reasonably combined and matched by those skilled in the art to form a variety of equivalent solutions without departing from or changing the design concept and technical effects of the present invention. Therefore, the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM, characterized in that, The method includes the following steps: Step S1: Combine the load data collected by the power consumption data acquisition system of the distribution network and the meteorological data collected by the meteorological station near the distribution transformer area to form an original data set, and divide the original data set into 80% training data and 20% test data; Step S2: For the load data in the training data, use the Adaptive Chirp Mode Decomposition (ACMD) method to decompose the historical load sequence data, and obtain q subsequences; combine the meteorological data characteristics in the test data set; Step S3: Build a Bidirectional Long Short-Term Memory (BiLSTM) prediction model for each of the decomposed q subsequences; use the Sparrow Search Algorithm (SSA) to optimize the q BiLSTM models constructed, optimize the hyperparameter values of the models, and build an optimal prediction model according to the optimized values; Step S4: Predict each subsequence according to the established optimal prediction model, superimpose and combine the prediction results to obtain the final prediction result, compare with other prediction methods, and evaluate the prediction result through relevant evaluation indicators to analyze the prediction effect of the model.

2. The short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM according to claim 1, characterized in that, The ACMD method uses the framework of a recursive algorithm for matching pursuit to process multi-modal signals with strong time-varying characteristics; use ACMD to process the historical power signals required for load forecasting, and the specific formula is as follows: Where Q is the number of signal modes, Aq(t), fq(t), and These represent the instantaneous amplitude, instantaneous frequency, and initial phase of the q-th signal mode, respectively.

3. The short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM according to claim 2, characterized in that, Through the modulation and demodulation steps, rewrite Equation (1) as follows: Where bq(t) and dq(t) are demodulation operators. This represents the demodulation frequency DF. This is the instantaneous amplitude; it can be observed that when the demodulation frequency equals the instantaneous frequency, that is, when... When the frequency modulation term in the demodulated signal is 0, a pure amplitude modulation signal with the narrowest bandwidth is obtained.

4. The short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM according to claim 3, characterized in that, By minimizing the bandwidth of the demodulated signal, the ACMD algorithm performs intermediate frequency estimation, decomposes the signal, and estimates each signal component one by one. This optimization problem is solved by alternately updating the demodulated signal and the frequency function to achieve the decomposition of the original signal one by one. After obtaining the first signal component, subtract it from the original signal, and then use the remaining component as the new initial signal to continue the decomposition to obtain the second signal component. Repeat the cycle update until all signal components are obtained, as follows: Where ||...||2 is the L2 norm, used to represent the distance; " is the second derivative, whose physical meaning is the smoothness of the single-mode curve after decomposition; λ is the penalty factor; and The bandwidth is represented by N. For a discrete signal in a time series t = t0, ..., tN-1, with N sampling points, the discrete expression of the above formula is: Where s=[s(t0),…,s(t) N-1 )] T u q =[(b q ) T ,(d q ) T ] T b q =[b q (t0),…,b q (t N-1 )] T ,d q =[d q (t0),…,d q (t N-1 )] T Λ=diag[Ω,Ω] is a block diagonal matrix, Ω is a second-order difference matrix, and F q This is the reference frequency used for demodulation.

5. The short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM according to claim 1, characterized in that, The BiLSTM is composed of two layers of LSTM combined, and at the same time, the two layers of LSTM are in forward and reverse sequence directions. The forward layer is used to capture the feature sequence before the current time node, and the reverse layer is used to capture the feature sequence from the future moment to the current time node.

6. The short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM according to claim 1, characterized in that, The process of establishing the prediction model seeks a better position by continuously updating the positions of various sparrows in the D-dimensional solution space and comparing the fitness values, that is, solving the optimal solution of the objective function. The various sparrows include discoverers, joiners, and alarmists.

7. A short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM according to claim 6, characterized in that, The calculation formula for updating the position of the discoverer is as follows: in, It represents the d-th dimension position of individual i in the t-th iteration of the population; ɑ is a random number between 0 and 1; iter max is the maximum number of iterations for the population; Q is a random number that follows a Gaussian distribution; R2 is also a random number between 0 and 1; ST is the alarm threshold between 0.5 and 1; when the value is greater than the alarm value, the position will be updated randomly; when the value is less than the alarm value, the position will converge at the current position.

8. The short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM according to claim 7, characterized in that, The calculation formula for updating the position of the joiner is as follows: in, This represents the current position of the worst-performing sparrow individual. The best individual position is represented by `rand{-1,1}`, which means randomly selecting -1 and 1. `n` is the total number of individuals. When `i>n / 2`, it means that a low-fitness entrant has not obtained food at the current position and needs to update its position. When i < n / 2, it means that it is currently near the optimal position, and continue to forage near the current position.

9. A short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM according to claim 8, characterized in that, The calculation formula for updating the position of the alarmist is as follows: Where β is a random number following a Gaussian distribution; f i f is the current individual fitness value. b f is the fitness value of the best individual in the current population. w ε is the worst individual fitness value; K is a random number between -1 and 1; ε is a very small non-zero number; if the alarm user is in the current optimal position, they will escape to the vicinity of the optimal position; if they are not in the current optimal position, they will escape to the vicinity of the optimal position.

10. A short-term load combination forecasting method based on adaptive chirped mode decomposition and SSA-BiLSTM according to any one of claims 1-9, characterized in that, Obtain the SSA-BiLSTM optimized prediction model through the above content. The establishment process includes: Step SA1: Set the number of sparrow populations of SSA, the proportion of sparrows with different task divisions, and the maximum number of iterations; Step SA2: Initialize the parameters to be optimized in the solution space; Step SA3: Use the corresponding parameter values to build a BiLSTM model to train the samples and calculate the fitness value; Step SA4: Update the sparrow population position based on the fitness value transmitted from BiLSTM to obtain a new solution. Transmit the corresponding solution, i.e. the parameter values, to BiLSTM for training. Step SA5: Repeat steps S2 and S3 until the iteration is complete; Step SA6: Output the optimal combination of parameter values ​​and establish the optimal BiLSTM prediction model.

Images