Two-stage network supply load prediction method considering optimal parameters

Abstract

The invention discloses a two-stage network supply load prediction method considering optimal parameters, and the method comprises the steps: defining an integrated empirical mode decomposition (EEMD) evaluation index suitable for load prediction, and solving a decomposition parameter optimal solution under the minimum evaluation index through employing a flower pollination algorithm FPA; performing hyper-parameter optimization on the BiGRU by using a flower pollination algorithm (FPA), and establishing a prediction model of each modal component to obtain a preliminary prediction result of each modal component; and superposing the preliminary prediction result and the actual value of each modal component to perform subtraction to obtain an error sequence as an output, combining each modal component with a short-term factor as an input, and establishing a mapping relation between the error sequence and the input, thereby performing error correction to obtain a final load prediction result. According to the two-stage network supply load prediction method considering the optimal parameters, high-precision load prediction can be carried out on a load sequence with strong nonlinearity and high volatility.

Description

Technical Field

[0001] The invention relates to the technical field of power transmission network load prediction, and in particular to a two-stage power transmission network load prediction method taking into account optimal parameters. Background Art

[0002] With the large-scale access of clean energy such as small hydropower, photovoltaic and wind power, the load sequence in areas with many small hydropower stations has strong nonlinearity and high complexity, which greatly increases the difficulty of load forecasting, which is not conducive to grid dispatching and clean energy consumption. Therefore, improving the accuracy of grid load forecasting is crucial to enhancing the grid dispatching capability, power supply reliability and economy.

[0003] At present, in order to deal with the nonlinearity and volatility of load series, deep learning methods represented by deep neural networks are mainly selected, which have advantages in processing complex nonlinear load data and improving load forecasting accuracy. In order to deal with the impact of load series volatility and nonlinearity, the combined model method based on time series decomposition is mainly used, but the classic signal decomposition method still has the following problems:

[0004] (1) In the actual decomposition process, there is a lack of guiding standards for parameter setting. The parameter setting lacks theoretical support and has strong randomness, which often leads to unsatisfactory decomposition results.

[0005] (2) The existence of residual components can easily cause prediction errors and reduce prediction accuracy. In the current load forecasting field, although the combined model has a good performance in load forecasting, it does not take into account the impact of the model's optimal parameters and various short-term factors such as weather, day type, etc., as well as the model's inherent errors on the prediction results.

[0006] Therefore, there is an urgent need for a network load forecasting method that can determine the optimal parameters for each prediction stage and reduce the inherent errors of the model and the impact of short-term factors such as weather and day type on prediction accuracy, in order to maximize the prediction accuracy in the case of strong randomness and numerous influencing factors in the power supply and load forecast of small hydropower networks. Summary of the invention

[0007] The present invention provides a two-stage network load forecasting method taking into account optimal parameters, which can perform high-precision load forecasting for load sequences with strong nonlinearity and high volatility.

[0008] The technical solution adopted by the present invention is:

[0009] The two-stage grid load forecasting method taking into account the optimal parameters comprises the following steps:

[0010] Step 1: Define the integrated empirical mode decomposition (EEMD) evaluation index suitable for load forecasting, and use the flower pollination algorithm (FPA) to solve the optimal solution of the decomposition parameters under the minimum evaluation index;

[0011] Step 2: Use the flower pollination algorithm (FPA) again to optimize the hyperparameters of the bidirectional gated recurrent unit (BiGRU), establish the prediction model of each modal component, and obtain the preliminary prediction results of each modal component;

[0012] Step 3: Superimpose the preliminary prediction results of each modal component and the actual value to obtain the error sequence as the output. Each modal component is combined with the short-term factor as the input to establish a mapping relationship between the two, so as to perform error correction and obtain the final load forecast result.

[0013] In step 1, first, define the decomposition loss L in the decomposition process loss As an evaluation index; secondly, the flower pollination algorithm (FPA) is used to find the decomposition loss L loss The minimum number of operations M and the standard deviation of white noise A are obtained, and the original load data are decomposed using integrated empirical mode decomposition (EEMD).

[0014] The Flower Pollination Algorithm (FPA) optimizes the ensemble empirical mode decomposition (EEMD) decomposition, which specifically includes the following steps:

[0015] Step 1.1: First define the EEMD algorithm as a function f eemd , the decomposition and reconstruction process is expressed as:

[0016] {y m (t)m=1,2,…,J,t=1,2,…,T}

[0017] =f eemd (f(t),M,A) (1);

[0018]

[0019] Where: f(t) is the original signal; f'(t) is the reconstructed signal; y m (t) is the mth modal component IMF decomposed by EEMD; T is the time length; m represents the modal component number; J is the number of modal components; M represents the number of operations; A represents the standard deviation of white noise; f eemd f(t),M,A) represents eemd The independent variables of the function are the decomposed signal, the number of operations and the standard deviation of white noise; f'(t) represents the superposition of all modal components, that is, the reconstructed signal.

[0020] Step 1.2: Based on the conventional evaluation indicators in the field of time series prediction, a decomposition loss L is proposedloss Evaluation criteria:

[0021]

[0022] Where: L loss The actual meaning of is the mean absolute error between the reconstructed signal and the original signal. The smaller this value is, the more information the mode contains.

[0023] Step 1.3: Determine the initial parameters of the Flower Pollination Algorithm (FPA). Among them, the conversion probability is generally a fixed value and cannot be dynamically adjusted according to the number of iterations, which will affect the search for the global optimal solution to a certain extent. Therefore, to improve the search ability of the algorithm, an adaptive conversion probability is introduced:

[0024]

[0025] Where: p is the conversion probability, p max and p min are the maximum and minimum values of the conversion probability respectively, taking 0.9 and 0.2; Max_iter represents the maximum number of iterations, a represents the current number of iterations; Max_iter - a represents the maximum number of iterations minus the current number of iterations;

[0026] It can be seen from Equation (4) that in the initial stage of the algorithm, the value of a is small and p is large, and the algorithm is more inclined to global search; in the later stage of the algorithm, p is small and it is more inclined to local search. Therefore, introducing an adaptive conversion probability is beneficial to balancing global search and local search and improving the global optimization ability of the algorithm;

[0027] Step 1.4: According to the number of optimal parameters to be solved, set the optimization dimension to 2, that is, the flower individual contains 2 values, which represent the number of operations M and the standard deviation A of the added white noise respectively;

[0028] Step 1.5: Generate an initial population with a uniform distribution and calculate the fitness value according to Equation (3). The fitness function is the function defined in Step 1.2, that is, Equation (3)), and then retain the global optimal individual and the optimal fitness of the current population.

[0029] Step 1.6: Compare the size of the random number ε and the conversion probability p. If ε < p, then choose cross-pollination to generate a new individual, as shown in Equation (5); otherwise choose self-pollination to generate a new individual, as shown in Equation (6):

[0030]

[0031] Where: ε is a random number obeying a uniform distribution, taking values between (0, 1), is the flower individual i at time t; is the updated flower individual i at time t + 1; is another flower individual j of the same population at time t: g * is the global optimal individual; γ is the scaling factor for controlling the step length; L is the levy flight formula.

[0032] Step 1.7: Considering that the number of operations M is a positive integer, the discrete variables in the new individuals are processed by rounding to obtain approximate integer values ​​as new population individuals; the continuous variable A does not need to be processed and is directly replaced by the new individual. Specifically, the white noise standard deviation in the fitness is processed. After comparing the fitness, the white noise standard deviation under the smaller fitness value is replaced by the compared one.

[0033] Step 1.8: Perform an interval check on the new population. If there are individuals within the interval, reset the interval based on the individual values, i.e., set the maximum and minimum individuals outside the interval as the upper and lower limits of the interval.

[0034] Step 1.9: Calculate the fitness value of the new population according to formula (3). The fitness function is the function defined in step 1.2, that is, formula (3), and update the global optimal individual and the optimal fitness minL loss , that is, the parameter with the smallest loss is the optimal parameter. When the iteration is completed, the global optimal individual and optimal fitness of the current population are output, that is, the number of operations under the minimum loss and the standard deviation of white noise are output. The global optimal individual is the optimal decomposition parameter of EEMD, and the decomposition loss is minimized at this time.

[0035] In step 1, by defining the decomposition loss evaluation index, the EEMD decomposition parameter solution problem is converted into a constrained optimization problem, and the formula is as follows:

[0036]

[0037] In the formula, st represents the inequality constraint based on the value range of M and A, and min represents obtaining M and A with the minimum decomposition loss.

[0038] In step 2, the flower pollination algorithm (FPA) is used to optimize the number of hidden layer neurons and learning rate parameters of the BiGRU neural network, and an FPA-BiGRU network model is built for each modal component, such as Figure 2 As shown, FPA is used to optimize hyperparameters for each modal component (IMF).

[0039] The Flower Pollination Algorithm (FPA) optimization of the BiGRU neural network includes the following steps:

[0040] Step 2.1: Generate the initial population of flowers:

[0041] Determine the value range of the optimization parameter learning rate and hidden layer neurons. Each pollen gamete in the population represents a combination of BiGRU neural network hyperparameters.

[0042] Step 2.2: Bring in the generated initial parameter combination and the historical data required for prediction, and input them into the BiGRU neural network for prediction;

[0043] Step 2.3: According to the prediction results, select the mean absolute percentage error (MAPE) as the fitness function and calculate the fitness value;

[0044]

[0045] In the formula, T is the number of samples, is the predicted value, x i is the true value.

[0046] The pollen gametes with the smallest mean absolute percentage error (MAPE) value in the current population are selected as the optimal combination for preservation, where the fitness calculation formula is as follows;

[0047] Step 2.4: Determine whether the flower pollination algorithm (FPA) meets the maximum number of iterations. If not, update the population and repeat steps 2.2 and 2.3;

[0048] Step 2.5: After reaching the maximum number of iterations, set the optimal solutions of all populations as the parameters of the BiGRU neural network to complete the model training.

[0049] In step 3, the error sequence is used as output, and each modal component is combined with short-term factors including weather and day type as input, a mapping relationship between the two is established, and an FPA-BiGRU error correction model is constructed, so as to perform error correction and obtain the final load forecast result.

[0050] The mapping relationship between input and output:

[0051] P 0 (t) = {P'(t), H(t)} (8);

[0052] Where P 0 (t) represents the error sequence prediction result, P'(t) is the error sequence of power prediction, and H(t) represents the meteorological data.

[0053] The FPA-BiGRU error correction model subtracts the original load data from the first-stage prediction value, combines the meteorological data as the error model input, optimizes the BiGRU model parameters through the FPA algorithm, and obtains the prediction results of the error sequence.

[0054] The error correction value is the prediction result of the second stage, P 0(t) = {P'(t), H(t)}.

[0055] The present invention provides a two-stage grid load forecasting method taking into account optimal parameters, and the technical effects are as follows:

[0056] 1) The present invention proposes a two-stage network load forecasting method taking into account the optimal parameters. The proposed method uses the FPA algorithm to optimize the parameters of EEMD decomposition and BiGRU neural network to maximize the prediction accuracy, and the two-stage prediction process fully considers the impact of the model inherent error and short-term factors on the prediction results, and finally constructs a two-stage network load forecasting model taking into account the optimal parameters for prediction.

[0057] 2) The FPA optimized EEMD strategy proposed in the present invention can effectively reduce the decomposition loss and improve the decomposition effect. Through comparative analysis, the strategy greatly improves the accuracy of the subsequent prediction model by efficiently and completely extracting effective information from the data.

[0058] 3) The present invention uses the FPA algorithm to optimize hyperparameters for different modal components, and specifically establishes a component prediction model based on FPA-BiGRU to predict the prediction values ​​of each modal component. By comparing with multiple groups of models, it is proved that FPA-BiGRU can effectively complete the task of modal component prediction while ensuring accuracy and having extremely fast speed, demonstrating the ability of FPA-BiGRU in grid load prediction.

[0059] 4) Error correction experiments are conducted on the basis of maximum consideration of the optimal parameters. By comparing the prediction results of three models: model 1: no error correction prediction, model 2: the error correction model only uses the modal components obtained by decomposition as input, and model 3: the error correction model uses modal components and short-term factors such as meteorology and day type as input, it is successfully verified that full consideration of error correction and short-term factors is conducive to improving prediction accuracy.

[0060] 5) Through comparative studies of multiple groups of prediction models, the MAPE, RMSE, and MAE prediction errors of the model proposed in the present invention are all smaller than those of the comparison model, which proves that this method has superior prediction accuracy in grid load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0062] Figure 1 This is the flow chart of the FPA-BiGRU algorithm.

[0063] Figure 2 Figure 2 is a schematic diagram of the two-stage load prediction process.

[0064] Figure 3 Schematic diagram of the iterative process of decomposing losses for different algorithms.

[0065] Figure 4 is the FPA-EEMD load data decomposition result.

[0066] Figure 5 Neural network prediction results for different decomposition parameter settings.

[0067] Figure 6 Fitness curves of BiGRU optimized for different algorithms.

[0068] Figure 7 The prediction results of FPA, PSO and SSA combined with BiGRU.

[0069] Figure 8 This is the load forecast result taking into account error correction. DETAILED DESCRIPTION

[0070] A two-stage grid load forecasting method considering optimal parameters is proposed. In the first stage, an ensemble empirical mode decomposition (EEMD) evaluation index is proposed to measure the decomposition loss, and the flower pollination algorithm (FPA) is used to find the optimal combination of decomposition parameters under this standard, so as to reduce the decomposition loss and improve the decomposition effect. Then, the BiGRU neural network parameters are optimized by FPA, and a multi-level BiGRU network model is constructed for each modal variable to increase the model complexity and improve the prediction accuracy. In the second stage, the difference between the actual value and the predicted value in the above stage is used as the model output, and each modal component and the short-term influencing factors affecting the grid load are used as input. The mapping relationship between the two is established through the FPA-BiGRU model, and the error correction model is constructed to reduce the inherent error of the model by mining the hidden information of the error sequence. This prediction method can perform high-precision load forecasting for the load characteristics of areas with many small hydropower stations.

[0071] The two-stage grid load forecasting process includes the following steps:

[0072] Step S1, first, with the goal of reducing decomposition loss, an integrated empirical mode decomposition (EEMD) evaluation index suitable for load forecasting is defined; then the flower pollination algorithm (FPA) is used to solve the optimal solution of the decomposition parameters under the minimum evaluation index;

[0073] Step S2: In order to maximize the prediction accuracy, the FPA algorithm is used again to optimize the hyperparameters of the bidirectional gated recurrent unit (BiGRU), and a prediction model for each modal component is established to obtain a preliminary prediction result for each modal component.

[0074] Step S3: Since the model error is mainly affected by the EEMD decomposition algorithm and short-term factors affecting the load sequence itself, the error sequence is obtained by superimposing the preliminary prediction results of each modal component and the actual value and subtracting them as the output. Each modal component is combined with short-term factors such as weather and day type as input to establish a mapping relationship between the two, thereby performing error correction and obtaining the final load forecast result.

[0075] In step S1, the decomposition loss L in the decomposition process is first defined as loss As the evaluation index, FPA is used to find the loss The optimal number of operations M and the white noise standard deviation A under the minimum condition are obtained, so that the original load data of EEMD is decomposed and the FPA algorithm optimizes the EEMD decomposition. The specific steps are as follows:

[0076] Step 1.1, first define the EEMD algorithm as a function f eemd , the decomposition and reconstruction process is expressed as:

[0077] {y m (t)|m=1,2,…,J,t=1,2,…,T}

[0078] =f eemd (f(t),M,A) (1);

[0079]

[0080] Where: f(t) is the original signal, f'(t) is the reconstructed signal, y m (t) is the mth modal component (IMF) decomposed by EEMD; T is the time length, m represents the modal component number, and J is the number of modal components.

[0081] Step 1.2: Based on the conventional evaluation indicators in the field of time series prediction, a decomposition loss evaluation standard L is proposed loss , the formula is:

[0082]

[0083] Where: L loss The actual meaning of is the mean absolute error between the reconstructed signal and the original signal. The smaller this value is, the more information the mode contains.

[0084] Step 1.3, determine the initial parameters of the FPA algorithm, where the conversion probability is generally a fixed value and cannot be dynamically adjusted according to the number of iterations, which will affect the search for the global optimal solution to a certain extent. Therefore, in order to improve the algorithm search capability, an adaptive conversion probability is introduced, such as the formula:

[0085]

[0086] where p is the conversion probability, p max and p min are the maximum and minimum values of the conversion probability, taking 0.9 and 0.2 respectively; Max_iter represents the maximum number of iterations, and t represents the current number of iterations. It can be seen from Equation (4) that in the initial stage of the algorithm, the value of t is small and p is large, and the algorithm is more inclined to global search; in the later stage of the algorithm, p is small and it is more inclined to local search. Therefore, introducing an adaptive conversion probability is beneficial to balancing global search and local search and improving the global optimization ability of the algorithm.

[0087] Step 1.4: According to the number of optimal parameters to be solved, set the optimization dimension to 2, that is, the flower individual contains 2 values, which respectively represent the number of operations M and the standard deviation A of the added white noise.

[0088] Step 1.5: Generate an initial population with a uniform distribution, calculate the fitness value, and then retain the global optimal individual and the optimal fitness of the current population.

[0089] Step 1.6: Compare the size of the random number ε and the conversion probability p. If ε < p, then select cross-pollination to generate a new individual, as shown in Equation (5). Otherwise, select self-pollination to generate a new individual, as shown in Equation (6).

[0090]

[0091] where: ε is a random number following a uniform distribution, taking values between (0, 1), is the flower individual i at time t; is the updated flower individual i at time t + 1; is another flower individual j in the same population at time t: g * is the global optimal individual; γ is a scaling factor for controlling the step size; L is the levy flight formula.

[0092] Step 1.7: Considering that the number of operations M is a positive integer, therefore, in accordance with the rounding method, process the discrete variables in the new individual to obtain an approximate integer value as the individual of the new population; for the continuous variable A, no processing is required and it is directly replaced with the new individual.

[0093] Step 1.8: Conduct an interval check on the new population. If there are individuals within the interval range, then reset the interval range according to the individual values, that is, set the maximum and minimum individuals outside the interval as the upper and lower limits of the interval.

[0094] Step 1.9: Calculate the fitness value of the new population and update the global optimal individual and the optimal fitness. When the iteration is completed, output the global optimal individual and the optimal fitness of the current population, where the global optimal individual is the optimal decomposition parameter of EEMD, and at this time the decomposition loss is the smallest.

[0095] In step S1, by defining the decomposition loss evaluation index, the EEMD decomposition parameter solution problem is converted into a constrained optimization problem, and the formula is as follows:

[0096]

[0097] In the formula, st represents the inequality constraint based on the value range of M and A, and min represents obtaining M and A with the minimum decomposition loss.

[0098] In step S2, the number of hidden layer neurons and the learning rate parameters of BiGRU are optimized using the FPA algorithm, and an FPA-BiGRU network model is built for each modal component to construct a component prediction model. The FPA algorithm optimizes the BiGRU neural network and includes the following steps:

[0099] Step S2.1: Generate an initial population of flowers. Determine the value range of optimization parameters (learning rate and hidden layer neurons). Each pollen gamete in the population represents a combination of BiGRU hyperparameters.

[0100] Step S2.2: Bring in the generated initial parameter combination and the historical data required for prediction into the BiGRU neural network for prediction;

[0101] Step S2.3: According to the prediction results, the fitness function, namely the mean absolute percentage error (MAPE), is calculated, and the pollen gametes with the smallest MAPE value in the current population are selected as the optimal combination for preservation.

[0102] Step S2.4: Determine whether FPA meets the maximum number of iterations. If not, update the population and repeat the above steps.

[0103] Step S2.5: After reaching the maximum number of iterations, the optimal solutions of all populations are set as the parameters of BiGRU to complete the model training.

[0104] In step S3, the error sequence is obtained by subtracting the initial prediction result of the superimposed modal component and the actual value as the output, and each modal component is combined with short-term factors such as weather and day type as input to establish a mapping relationship between the two, and construct an FPA-BiGRU error correction model to perform error correction and obtain the final load forecast result. The two-stage network load forecasting includes the following steps:

[0105] Step S3.1: First define the decomposition loss L in the decomposition process loss As the evaluation index, FPA is used to find the lossThe number of operations and the standard deviation of white noise under the minimum condition are obtained, so that the original load data of EEMD is decomposed to reduce its complexity. At the same time, the FPA algorithm is used to optimize the number of hidden layer neurons and learning rate parameters of BiGRU. The FPA-BiGRU network model is built for each modal component, and the component prediction model is constructed to predict each modal component. Finally, the predicted values ​​of each component are superimposed to obtain the predicted load sequence.

[0106] Step S3.2: The error sequence is obtained by subtracting the predicted load sequence of the first stage from the actual value as the output. Each modal component is combined with short-term factors such as weather and day type as the input of the second stage. A mapping relationship between the two is established, and the FPA-BiGRU error correction model is constructed to perform error correction and obtain the final load forecast result.

[0107] Example:

[0108] Considering the relevant influencing factors of the load of the small hydropower grid, the data collected in this paper include load data, 7 common meteorological and daily types, among which meteorological data include rainfall, maximum temperature, minimum temperature, average temperature, wind speed, humidity and air pressure. The data set comes from a prefecture-level city in the southwest region from January 1, year Z to February 18, year (Z+3), with an interval of 1h. The total data set is divided according to the ratio of 8:2, and 916 days of data are selected as the training set and 229 days of data are selected as the test set. The installed capacity of small hydropower in this prefecture-level city is 950MW, the maximum power load capacity is 3500MW, and the average power load capacity is 2150MW.

[0109] The superiority of FPA-EEMD is verified by comparing the loss size in three cases, namely, the default parameters, the particle swarm algorithm (PSO) optimized EEMD and the FPA optimized EEMD. The iterative process of decomposing the loss is as follows Figure 3 As shown in Figure 1, the initial parameters of PSO and FPA are: the number of individuals in the population is 20, and the maximum number of iterations is 100. Among them, the acceleration constant of the learning factor of the particle swarm algorithm is set to 1.4. The EEMD decomposition parameters and decomposition losses of the three cases are shown in Table 1.

[0110] Table 1 EEMD parameter settings and decomposition losses

[0111]

[0112] In summary, finding the optimal parameters of EEMD through the FPA algorithm and the PSO algorithm can effectively reduce the decomposition loss, and the decomposition loss of FPA-EEMD is the smallest, indicating that FPA can effectively find the optimal parameters. The decomposition sequence of FPA-EEMD is as follows Figure 4As shown in Figure 1, the original load data is decomposed by EEMD to obtain 12 components. The other methods will not be described in detail.

[0113] By Figure 5 It can be seen that by comparing the prediction results with and without the EEMD algorithm, the MAE and RMSE of BiGRU are 290.21MW and 330.88MW respectively, which are much higher than the prediction results of the EEMD algorithm, indicating the effectiveness of the signal decomposition algorithm in improving the prediction accuracy; and after finding the optimal parameter solution through PSO and FPA, the prediction results are further improved, indicating the necessity of EEMD parameter optimization; and FPA-EEMD-BiGRU has the smallest error, with MAE and RMSE of 150.2MW and 189.11MW respectively. In summary, the model established by combining FPA-EEMD with the BiGRU neural network has higher prediction accuracy.

[0114] The parameter settings of the FPA algorithm and the optimization range of the BiGRU neural network related parameters are shown in Table 2 and Table 3 respectively. The optimal parameters of BiGRU obtained by FPA optimization are {[20,25], 0.002}.

[0115] Similarly, the commonly used particle swarm optimization (PSO) and grey wolf optimization (GWO) are selected for comparative analysis with FPA. The iteration curves are shown in Figure 6 shown.

[0116] Table 2 FPA algorithm parameter settings

[0117]

[0118] Table 3Bi GRU parameter optimization range

[0119]

[0120] Depend on Figure 6 It can be seen that among the three optimization algorithms, the convergence speed and convergence accuracy of the FPA algorithm are improved compared with other optimization algorithms, and convergence is achieved after 20 iterations. The fitness value of FPA-BiGRU is 0.02.

[0121] In order to verify the necessity of optimizing the prediction model parameters, the prediction results of the four models of FPA-BiGRU, PSO-BiGRU, SSA-BiGRU and BiGRU are compared. Figure 7 And as shown in Table 4.

[0122] Table 4 Comparison of results of different neural networks

[0123]

[0124] From the above indicators, FPA-BiGRU, PSO-BiGRU and SSA-BiGRU are all lower than BiGRU, indicating the effectiveness of optimizing the prediction model parameters; and compared with the three optimization algorithms, FPA has the smallest error, with MAE and RMSE of 141.49MW and 179.14MW respectively, indicating that the FPA algorithm has better optimization effect than the other two algorithms. In summary, the combination of FPA and BiGRU neural network has higher accuracy.

[0125] In summary, the FPA algorithm combined with EEMD and BiGRU can improve the prediction accuracy. Therefore, this paper combines the two cases, finds the optimal parameters of the EEMD and BiGRU neural networks respectively through FPA, and conducts error correction experiments on the basis of maximizing the consideration of the optimal parameters. By comparing the prediction results of three modes: mode 1 (model1): no error correction prediction, mode 2 (model2): the error correction model only uses the modal components obtained by decomposition as input, mode 3 (model3): the error correction model uses the modal components and short-term factors such as weather and day type as input, verify the impact of error correction and short-term factors on the prediction results, such as Figure 8 And as shown in Table 5.

[0126] Table 5 Comparison of results of different modes

[0127]

[0128] Depend on Figure 8 As shown in Table 5, the use of FPA for comprehensive optimal parameter solution can improve the prediction accuracy to a certain extent, and R2 reaches 0.85, which has a fairly high interpretability, but is lower than the prediction accuracy of the error correction model. As shown in Table 5, the evaluation indicators RMSE, MAPE and MAE of model2 and model3 have all decreased, among which model3 has the highest prediction accuracy, indicating that factors such as meteorology and day type can improve the effect of error correction, and the determination coefficient R2 of model3 has also reached 0.95, indicating that its prediction results are also more in line with the true value.

[0129] The present invention provides a two-stage grid load forecasting method taking into account optimal parameters. Aiming at the strong randomness and numerous influencing factors of grid load forecasting in areas with many small hydropower stations, in order to determine the optimal parameters for each prediction stage and reduce the inherent error of the model and the influence of short-term factors such as weather and day type on the prediction accuracy, and to maximize the prediction accuracy, a two-stage small hydropower grid power supply load forecasting method taking into account optimal parameters is proposed according to the idea of ​​"two-stage grid load forecasting". The first stage: First, with the goal of reducing decomposition loss, an EEMD evaluation index suitable for load forecasting is defined; then FPA is used to solve the optimal solution of the decomposition parameters under the minimum evaluation index. In order to maximize the prediction accuracy, the FPA algorithm is used again to optimize the hyperparameters of BiGRU, establish the prediction model of each modal component, and obtain the preliminary prediction results of each modal component; the second stage: since the model error is mainly affected by the EEMD decomposition algorithm and the short-term factors affecting the load sequence itself, the initial prediction value and the actual value of the first stage are subtracted to obtain the error sequence as the output, and each modal component is combined with short-term factors such as weather and day type as the input of the second stage, and a mapping relationship between the two is established, so as to perform error correction and obtain the final load forecasting result.

Claims

1. A two-stage grid load forecasting method taking into account optimal parameters, characterized in that It includes the following steps: Step 1: Define the evaluation index of the Ensemble Empirical Mode Decomposition (EEMD) applicable to load forecasting, and use the Flower Pollination Algorithm (FPA) to solve the optimal solution of the decomposition parameters under the minimum evaluation index. Step 2: Use the Flower Pollination Algorithm (FPA) again to optimize the hyperparameters of the Bidirectional Gated Recurrent Unit (BiGRU), establish prediction models for each modal component, and obtain the preliminary prediction results of each modal component. Step 3: Subtract the sum of the preliminary prediction results of each modal component from the actual value to obtain an error sequence as the output. Combine each modal component with short-term factors as the input, establish the mapping relationship between the two, and thus perform error correction to obtain the final load prediction result.

2. The two-stage grid load forecasting method taking into account the optimal parameters according to claim 1 is characterized in that: In step 1, first, define the decomposition loss L in the decomposition process loss As an evaluation index; secondly, the flower pollination algorithm FPA is used to find the decomposition loss L loss The minimum number of operations M and the white noise standard deviation A are obtained, and the original load data are decomposed using the integrated empirical mode decomposition EEMD.

3. The two-stage grid load forecasting method taking into account the optimal parameters according to claim 2 is characterized by: The Flower Pollination Algorithm (FPA) optimizes the Ensemble Empirical Mode Decomposition (EEMD) decomposition, specifically including the following steps: Step 1.1: First define the EEMD algorithm as a function f eemd , the decomposition and reconstruction process is expressed as: {y m (t)m=1,2,…,J,t=1,2,…,T} =f eemd (f(t),M,A) (1); Where: f(t) is the original signal; f'(t) is the reconstructed signal; y m (t) is the mth modal component IMF decomposed by EEMD; T is the time length; m represents the modal component number; J is the number of modal components; M represents the number of operations; A represents the standard deviation of white noise; f eemd f(t),M,A) represents eemd The independent variables of the function are the decomposed signal, the number of operations and the standard deviation of white noise; f'(t) represents the superposition of all modal components, i.e. the reconstructed signal; Step 1.2: Based on the conventional evaluation indicators in the field of time series prediction, a decomposition loss L is proposed loss Evaluation criteria: Where: L loss The actual meaning of is the mean absolute error between the reconstructed signal and the original signal. The smaller this value is, the more information the mode contains. Step 1.3: Determine the initial parameters of the Flower Pollination Algorithm (FPA), and introduce the adaptive transition probability: Where: p is the conversion probability, p max and p min Divided into the maximum and minimum values ​​of the conversion probability; Max_iter represents the maximum number of iterations, a represents the current number of iterations; Max_iter-a represents the maximum number of iterations minus the current number of iterations; As can be seen from Equation (4), at the initial stage of the algorithm, the value of a is small and p is large, and the algorithm is more inclined to global search; at the later stage of the algorithm, p is small and it is more inclined to local search. Step 1.4: According to the number of optimal parameters to be solved, set the optimization dimension to 2, that is, each flower individual contains 2 values, which represent the number of operations M and the standard deviation A of the added white noise respectively. Step 1.5: Generate an initial population with a uniform distribution, calculate the fitness value according to Equation (3), and then retain the global optimal individual and the optimal fitness of the current population. Step 1.6: Compare the sizes of the random number ε and the transition probability p. If ε < p, then cross-pollination is selected to generate new individuals, as shown in Equation (5); otherwise, self-pollination is selected to generate new individuals, as shown in Equation (6). Where: ε is a random number that follows a uniform distribution and takes a value between (0, 1), is the flower individual i at time t; is the flower individual i updated at time t+1; is another flower individual j of the same population at time t: g * is the global optimal individual; γ is the scaling factor for controlling the step length; L is the levy flight formula; Step 1.7: Considering that the number of operations M is a positive integer, process the discrete variables in the new individual to obtain an approximate integer value as the individual of the new population; for the continuous variable A, there is no need to process it, and it is directly replaced with the new individual. Step 1.8: Conduct an interval check on the new population. If there are individuals within the interval range, re-set the interval range according to the individual values, that is: set the maximum and minimum individuals outside the interval as the upper and lower limits of the interval. Step 1.9: Calculate the fitness value of the new population according to formula (3), and update the global optimal individual and optimal fitness minL loss , that is, the parameter with the smallest loss is the optimal parameter; when the iteration is completed, the global optimal individual and optimal fitness of the current population are output, that is, the number of operations under the minimum loss and the standard deviation of white noise are output, where the global optimal individual is the optimal decomposition parameter of EEMD, and the decomposition loss is minimized at this time.

4. The two-stage grid load forecasting method taking into account the optimal parameters according to claim 3 is characterized by: In the above Step 1, by defining the decomposition loss evaluation index, the problem of solving the EEMD decomposition parameters is transformed into a constrained optimization problem, and the formula is as follows: In the formula, s.t. represents the inequality constraint according to the value ranges of M and A, and min represents obtaining M and A with the minimum decomposition loss.

5. The two-stage grid load forecasting method taking into account the optimal parameters according to claim 3 is characterized by: In the above Step 2, use the Flower Pollination Algorithm (FPA) to optimize the number of neurons in the hidden layer and the learning rate parameters of the BiGRU neural network, and build an FPA-BiGRU network model for each modal component respectively.

6. The two-stage grid load forecasting method taking into account the optimal parameters according to claim 5 is characterized in that: The Flower Pollination Algorithm (FPA) optimizes the BiGRU neural network, including the following steps: Step 2.1: Generate the initial population of flowers: Determine the value ranges of the optimization parameters learning rate and the number of neurons in the hidden layer. Each pollen gamete in the population represents a combination of BiGRU neural network hyperparameters. Step 2.2: Substitute the generated initial parameter combination and the historical data required for prediction into the BiGRU neural network for prediction. Step 2.3: According to the prediction results, select the Mean Absolute Percentage Error (MAPE) as the fitness function and calculate the fitness value. In the formula, T is the number of samples, is the predicted value, x i is the true value; Select the pollen gametes with the smallest mean absolute percentage error MAPE value in the current population as the optimal combination for preservation, where the fitness calculation formula is as follows; Step 2.4: Determine whether the flower pollination algorithm FPA meets the maximum number of iterations. If not, update the population and repeat steps 2.2 and 2.3; Step 2.5: After reaching the maximum number of iterations, set the optimal solutions of all populations as the parameters of the BiGRU neural network to complete the model training.

7. The two-stage grid load forecasting method taking into account the optimal parameters according to claim 1 is characterized by: In step 3, the error sequence is used as output, and each modal component is combined with short-term factors including weather and day type as input, a mapping relationship between the two is established, and an FPA-BiGRU error correction model is constructed, so as to perform error correction value and obtain the final load forecast result; The mapping relationship between input and output: P0(t)={P′(t),H(t)} (8); Where, P0(t) represents the error sequence prediction result, P'(t) represents the error sequence of power prediction, and H(t) represents meteorological data; The FPA-BiGRU error correction model subtracts the original load data from the first-stage prediction value, combines the meteorological data as the error model input, optimizes the BiGRU parameters through the FPA algorithm, and obtains the prediction results of the error sequence.

8. The two-stage grid load forecasting method taking into account the optimal parameters according to claim 7 is characterized in that: The error correction value is the prediction result of the second stage: P0(t)={P'(t),H(t)}.

Images