Method for predicting short-term load probability of power distribution network

Abstract

The invention specifically relates to a method for predicting the short-term load probability of a power distribution network, and belongs to the technical field of power distribution network load prediction. According to the method, the power distribution network load probability prediction is decomposed into two sub-problems of probability proportion selection and quantile prediction, the two sub-problems can be processed respectively, and the flexibility and efficiency of the power distribution network load probability prediction are improved. According to the method, a combined experience pool playback mechanism is selected to improve a double-delay depth deterministic strategy gradient algorithm, the network is updated by preferentially considering experience samples, high-value samples are particularly concerned, and the defect of a traditional TD3 algorithm experience playback mechanism is overcome, that is, random sampling of low-value samples can be possibly caused, and the probability that the high-value samples are randomly sampled is reduced. Therefore, the training effect and the final performance are influenced, and the accuracy of power distribution network load probability prediction is improved.

Description

Technical Field

[0001] The present invention specifically relates to a method for predicting short-term load probability of a distribution network, and belongs to the technical field of distribution network load prediction. Background Art

[0002] Short-term load forecasting of distribution networks is of great value in modern power systems. Through accurate load forecasting, grid operators can identify potential overload situations in advance, take timely measures to avoid power outages, optimize the allocation of power resources, ensure power supply stability and reliability, and reduce power losses and operating costs. Carrying out short-term load forecasting of distribution networks helps to balance supply and demand, fully cope with load uncertainty, and improve the flexibility and adaptability of the power grid.

[0003] Short-term load forecasting of distribution networks is of great value in modern power systems. Through accurate load forecasting, grid operators can identify potential overload situations in advance, take timely measures to avoid power outages, optimize the allocation of power resources, ensure power supply stability and reliability, and reduce power losses and operating costs. Carrying out short-term load forecasting of distribution networks helps to balance supply and demand, fully cope with load uncertainty, and improve the flexibility and adaptability of the power grid.

[0004] At present, the load forecasting methods of distribution network are divided into point forecasting and load probability forecasting. Compared with point forecasting, load probability forecasting has significant advantages. It can quantify load uncertainty, provide upper and lower boundaries, help power grid operators identify and manage potential risks, reduce power supply interruptions caused by load fluctuations, and provide a basis for decision support. A Chinese paper named "Short-term Load Forecasting Strategy Based on TD3 Variable Length Time Window Optimal Weighting" proposed a method of using twin delay deep deterministic policy gradient model (twin delay deep deterministic policy gradient, TD3) to predict the short-term load probability of distribution network. After reading, it was found that the existing technology uses TD3 algorithm to predict the short-term load probability of distribution network, which has the following disadvantages: the experience replay mechanism of TD3 algorithm may cause low-value samples to be frequently selected when randomly sampling experience samples, thereby affecting the training effect and final performance, resulting in poor prediction accuracy of the trained distribution network short-term load probability prediction model. Summary of the invention

[0005] The technical problem to be solved by the present invention is: how to improve the accuracy of short-term load probability prediction of distribution network.

[0006] In order to solve the above technical problems, the technical solution proposed by the present invention is: a method for predicting the short-term load probability of a distribution network, comprising the following steps:

[0007] Step 1: Establish the initial distribution network short-term load probability prediction model based on the quantile regression algorithm, as shown in the following formula (1):

[0008] (1)

[0009] In formula (1), is the quantile regression algorithm, x is the real load sequence of the distribution network input into the quantile regression algorithm; is the quantile probability ratio in the quantile regression algorithm; is the first regression coefficient; is the second regression coefficient;

[0010] Step 2: Define the time when the short-term load probability forecast of the distribution network is required as the load prediction time t, collect the n distribution network historical load data before the load prediction time t and form the distribution network historical load data set A, as shown in the following formula (1):

[0011] (2)

[0012] In formula (2), They are respectively the historical load data of the first distribution network before the load prediction time t to the historical load data of the nth distribution network;

[0013] Collect the distribution network real load data from the moment before the load prediction moment t to the moment before m and form a distribution network real load data set B, as shown in the following formula (3):

[0014] (3)

[0015] In formula (3), are respectively the actual load data of the distribution network from the moment before the load prediction moment t to the moment before m;

[0016] According to the sliding window method, all data in the distribution network historical load data set A are sliced ​​to form a distribution network historical load data time series sequence from the previous moment to the load prediction moment t to the previous m+1 moment, and all the distribution network historical load data time series sequences are collected to form a distribution network historical load data time series sequence set X, as shown in the following formula (4):

[0017] (4)

[0018] In formula (4), are respectively time series of historical load data of the distribution network from the moment before the load prediction moment t to the moment m+1 before; is the window length of the sliding window;

[0019] Step 3: Calculate the optimal quantile probability ratio of the initial distribution network short-term load probability forecasting model based on the joint experience pool replay-double-delay deep deterministic policy gradient algorithm , the optimal first regression coefficient and the optimal second regression coefficient , the specific steps are as follows:

[0020] Step 3.1: Establish the quantile probability ratio update model in the initial distribution network short-term load probability forecasting model as shown in the following formula (5):

[0021] (5)

[0022] In formula (5), is the updated quantile probability ratio; is the state action value function; x is the real load sequence of the distribution network in the input quantile probability proportional update model; is the quantile probability ratio to be updated in the input quantile probability ratio update model; is a neural network parameter in the initial distribution network short-term load probability prediction model; It is the action space; is the action index value in the action space; is the quantile probability; is the number of actions in the action space H;

[0023] The regression coefficient updating model in the initial distribution network short-term load probability prediction model is established as shown in the following formula (6):

[0024] (6)

[0025] In formula (6), is the updated regression coefficient, is the regression coefficient to be updated; is the gradient calculation symbol; is the empirical loss function; is the learning rate in the initial distribution network short-term load probability forecasting model; is the size of the sampling experience pool; is the correction parameter; is the number of experiences in the predictor buffer; is the empirical sampling probability; is the maximum importance sampling weight; It is the priority of sampling experience; is the degree of sampling experience priority; is the quantile loss; It is the real load data of the distribution network; is the predicted load probability boundary of the distribution network;

[0026] The neural network parameter updating model in the initial distribution network short-term load probability prediction model is established as shown in the following formula (7):

[0027] (7)

[0028] In formula (7), are the updated neural network parameters; are the neural network parameters to be updated; is the update function; is the reward function;

[0029] Step 3.2: Initialize the quantile probability ratio in the initial distribution network short-term load probability forecasting model and the neural network parameters The value of is taken as the quantile probability ratio of the load at the m+1 time before the time t to be predicted and neural network parameters Initialize the first regression coefficient in the initial distribution network short-term load probability prediction model and the second regression coefficient The value of is used as the first regression coefficient of the initial distribution network short-term load probability forecasting model at the load prediction time t before m+1 time and the second regression coefficient ;

[0030] Step 3.3: The quantile probability ratio of the initial distribution network short-term load probability forecasting model at the m+1 time before the load prediction time t is calculated. and neural network parameters Enter the quantile probability ratio to update the model, and then The quantile probability ratio update model also outputs the quantile probability ratio calculated in the initial distribution network short-term load probability forecasting model at the m moments before the load prediction time t. ; , , and stated Substitute into the initial distribution network short-term load probability forecasting model to calculate the distribution network initial load probability upper boundary at the load prediction time t m before the time t and the lower bound of the initial load probability ;

[0031] Step 3.4: , , , and Substitute the regression update model into the calculation to obtain the first regression coefficient of the initial distribution network short-term load probability forecasting model at the load prediction time t before the first m time and the second regression coefficient

[0032] Step 3.5: , , , , , and Substitute into the neural network parameter update model to calculate the neural network parameters of the initial distribution network short-term load probability forecasting model at the m moments before the load prediction time t ;

[0033] Step 3.6: According to the principles of steps 3.2 to 3.5, the quantile probability ratio, the first regression coefficient, the second regression coefficient and the neural network parameters of the initial distribution network short-term load probability forecasting model from the m-1 moments before the load prediction moment t to the previous moment are calculated in sequence;

[0034] The quantile probability ratio of the initial distribution network short-term load probability forecasting model at the moment before the load prediction time t is calculated. and neural network parameters Enter the quantile probability ratio to update the model, and then The quantile probability ratio update model also outputs the quantile probability ratio of the initial distribution network short-term load probability forecasting model at the load prediction time t. ;

[0035] Select the quantile probability ratio of the initial distribution network short-term load probability forecasting model at the load prediction time t As the optimal quantile probability ratio of the initial distribution network short-term load probability forecasting model ; Select the first regression coefficient of the initial distribution network short-term load probability forecasting model at the moment before the load prediction time t and the second regression coefficient As the optimal first regression coefficient of the initial distribution network short-term load probability forecasting model and the optimal second regression coefficient ;

[0036] Step 4: The optimal quantile probability ratio obtained in step 3 , the optimal first regression coefficient and the optimal second regression coefficient Substituting into the initial distribution network short-term load probability prediction model forms the final distribution network short-term load probability prediction model, as shown in the following formula (8):

[0037] (8)

[0038] The Input the final distribution network short-term load probability forecasting model to calculate the upper boundary of the final load probability of the distribution network at the load prediction time t and the lower bound of the final load probability ; The upper limit of the final load probability and the lower bound of the final load probability The interval formed That is, the load probability interval of the distribution network at the load prediction time t.

[0039] Beneficial effects of the present invention: 1. The present invention decomposes the distribution network load probability prediction into two sub-problems, probability ratio selection and quantile prediction, which can be processed separately, thereby improving the flexibility and efficiency of the distribution network load probability prediction. 2. The present invention selects and adopts a joint experience pool replay mechanism to improve the double-delay deep deterministic policy gradient algorithm, and updates the network by giving priority to experience samples, paying special attention to high-value samples, thereby overcoming the drawbacks of the traditional TD3 algorithm experience replay mechanism, which may lead to random sampling of low-value samples, thereby affecting the training effect and final performance, and improving the accuracy of the distribution network load probability prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a prediction result diagram of a method for predicting short-term load probability of a distribution network according to the present invention. DETAILED DESCRIPTION

[0041] The following is a further description of a method for predicting short-term load probability of a distribution network according to the present invention in conjunction with the accompanying drawings and specific embodiments.

[0042] Example

[0043] The method for predicting the short-term load probability of the distribution network in this embodiment includes the following steps:

[0044] Step 1: Establish the initial distribution network short-term load probability prediction model based on the quantile regression algorithm, as shown in the following formula (1):

[0045] (1)

[0046] In formula (1), is the quantile regression algorithm, x is the real load sequence of the distribution network input into the quantile regression algorithm; is the quantile probability ratio in the quantile regression algorithm; is the first regression coefficient; is the second regression coefficient;

[0047] Step 2: Define the time when the short-term load probability forecast of the distribution network is required as the load prediction time t, collect the n distribution network historical load data before the load prediction time t and form the distribution network historical load data set A, as shown in the following formula (1):

[0048] (2)

[0049] In formula (2), They are respectively the historical load data of the first distribution network before the load prediction time t to the historical load data of the nth distribution network;

[0050] Collect the distribution network real load data from the moment before the load prediction moment t to the moment before m and form the distribution network real load data set B, as shown in the following formula (3):

[0051] (3)

[0052] In formula (3), They are the actual load data of the distribution network from the moment before the load prediction moment t to the moment before m;

[0053] According to the sliding window method, all data in the distribution network historical load data set A are sliced ​​to form a distribution network historical load data time series sequence from the moment before the load prediction time t to the moment before m+1, and all the distribution network historical load data time series sequences are collected to form a distribution network historical load data time series sequence set X, as shown in the following formula (4):

[0054] (4)

[0055] In formula (4), They are the time series of historical load data of the distribution network from the moment before the load prediction moment t to the moment m+1 before; is the window length of the sliding window;

[0056] Step 3: Calculate the optimal quantile probability ratio of the initial distribution network short-term load probability forecasting model based on the joint experience pool replay-double-delay deep deterministic policy gradient algorithm , the optimal first regression coefficient and the optimal second regression coefficient , the specific steps are as follows:

[0057] Step 3.1: Establish the quantile probability proportion update model in the initial distribution network short-term load probability forecasting model as shown in the following formula (5):

[0058] (5)

[0059] In formula (5), is the updated quantile probability ratio; is the state action value function; x is the real load sequence of the distribution network in the input quantile probability proportional update model; is the quantile probability ratio to be updated in the input quantile probability ratio update model; is the neural network parameter in the initial distribution network short-term load probability forecasting model; It is the action space; is the action index value in the action space; is the quantile probability; is the number of actions in the action space H;

[0060] The regression coefficient update model in the initial distribution network short-term load probability prediction model is established as shown in the following formula (6):

[0061] (6)

[0062] In formula (6), is the updated regression coefficient, is the regression coefficient to be updated; is the gradient calculation symbol; is the empirical loss function; is the learning rate in the initial distribution network short-term load probability forecasting model; is the size of the sampling experience pool; is the correction parameter; is the number of experiences in the predictor buffer; is the empirical sampling probability; is the maximum importance sampling weight; It is the priority of sampling experience; is the degree of sampling experience priority; is the quantile loss; It is the real load data of the distribution network; is the predicted load probability boundary of the distribution network;

[0063] The neural network parameter update model in the initial distribution network short-term load probability prediction model is established as shown in the following formula (7):

[0064] (7)

[0065] In formula (7), are the updated neural network parameters; are the neural network parameters to be updated; is the update function; is the reward function;

[0066] Step 3.2: Initialize the quantile probability ratio in the initial distribution network short-term load probability forecasting model and neural network parameters The value of is taken as the quantile probability ratio of the load at the m+1 moments before the load prediction time t and neural network parameters ; Initialize the first regression coefficient in the initial distribution network short-term load probability forecasting model and the second regression coefficient The value of is used as the first regression coefficient of the initial distribution network short-term load probability forecasting model at the m+1 time before the load prediction time t and the second regression coefficient ;

[0067] Step 3.3: The quantile probability ratio of the initial distribution network short-term load probability forecasting model at the m+1 time before the load prediction time t is calculated. and neural network parameters Enter the quantile probability ratio to update the model, and then The quantile probability ratio is also output in the update model to calculate the quantile probability ratio of the initial distribution network short-term load probability forecasting model at the m moments before the load prediction time t. ;Will , , and Substitute it into the initial distribution network short-term load probability forecasting model to calculate the upper boundary of the initial load probability of the distribution network at the m moments before the load prediction time t and the lower bound of the initial load probability ;

[0068] Step 3.4: , , , and Substitute into the regression update model to calculate the first regression coefficient of the initial distribution network short-term load probability forecasting model at the mth moment before the load prediction time t and the second regression coefficient

[0069] Step 3.5: , , , , , and Substitute it into the neural network parameter update model to calculate the neural network parameters of the initial distribution network short-term load probability forecasting model at the m moments before the load prediction time t ;

[0070] Step 3.6: According to the principles of steps 3.2 to 3.5, the quantile probability ratio, the first regression coefficient, the second regression coefficient and the neural network parameters of the initial distribution network short-term load probability forecasting model from the m-1 moments before the load prediction moment t to the previous moment are calculated in sequence;

[0071] The quantile probability ratio of the initial distribution network short-term load probability forecasting model at the moment before the load prediction time t is calculated. and neural network parameters Enter the quantile probability ratio to update the model, and then The quantile probability ratio is also output to update the model to calculate the quantile probability ratio of the initial distribution network short-term load probability forecasting model at the load prediction time t ;

[0072] Select the quantile probability ratio of the initial distribution network short-term load probability forecasting model at the load prediction time t Optimal quantile probability ratio as initial distribution network short-term load probability forecasting model ; Select the first regression coefficient of the initial distribution network short-term load probability forecasting model at the moment before the load prediction time t and the second regression coefficient As the optimal first regression coefficient of the initial distribution network short-term load probability forecasting model and the optimal second regression coefficient ;

[0073] Step 4: The optimal quantile probability ratio obtained in step 3 , the optimal first regression coefficient and the optimal second regression coefficient Substitute the initial distribution network short-term load probability prediction model into the final distribution network short-term load probability prediction model, as shown in the following formula (8):

[0074] (8)

[0075] Will Input the final distribution network short-term load probability forecasting model to calculate the upper boundary of the final load probability of the distribution network at the load prediction time t and the lower bound of the final load probability ;Upper bound of final load probability and the lower bound of the final load probability The interval formed That is, the load probability interval of the distribution network at the load prediction time t.

[0076] This example predicts the probability of distribution network load in the next 72 hours based on the historical load data and actual load data of a city's distribution network. The sample data of the historical load of the distribution network is shown in Table 1 below, and the sample data of the actual load of the distribution network is shown in Table 2 below.

[0077] t=1~8h (kW) t=9~16h (kW) t=17~24h (kW) 0 28.4708 41.4492 1.02954 25.4274 35.7365 1 29.2417 30.4048 4.08301 26.5788 25.8345 9.41845 31.5328 18.979 15.1332 36.8677 12.8848 19.3248 42.2021 15.5552 25.0401 44.1116 8.31929

[0078] Table 1

[0079] t=1~8h (kW) t=9~16h (kW) t=17~24h (kW) 3.70206 37.62 34.2138 0.2735 35.3385 29.2625 9.41739 30.3889 15.1684 14.3703 38.392 22.7885 17.8004 42.965 9.07746 23.5163 46.0148 11.7489 29.6131 44.4963 6.03625 37.2348 39.1656 2.61302

[0080] Table 2

[0081] The prediction results are as follows Figure 1 shown.

[0082] The prediction method in the present invention is compared with the prediction interval optimization method based on the mixed integer programming model (MLMIP) and the lightweight quantile gradient boosting regression tree (Light QGBRT), and the Winkler score, average coverage deviation score and sharpness score are used as indicators. The Winkler score comprehensively considers the reliability and width of the prediction interval. The lower the score, the higher the quality of the prediction interval. The average coverage deviation score measures the reliability of the prediction interval. A value close to zero indicates that the prediction interval can accurately contain the actual value. The sharpness score measures the width of the prediction interval. The lower the score, the narrower the prediction interval and the more accurate the prediction. The comparison is shown in Table 3 below.

[0083] Winkler score Average coverage deviation score Sharpness Rating MLMIP 47.52 -12.54 17.36 Light QGBRT 34.28 4.86 29.64 This method 21.57 -1.95 14.98

[0084] Table 3

[0085] By comparing Table 3, it can be clearly seen that the prediction accuracy of the present invention is significantly better than that of the existing method.

Claims

1. A method for predicting short-term load probability of a distribution network, characterized by: The following steps are involved: Step 1: Establish the initial distribution network short-term load probability prediction model based on the quantile regression algorithm, as shown in the following formula (1): （1） In formula (1), is the quantile regression algorithm, x is the real load sequence of the distribution network input into the quantile regression algorithm; is the quantile probability ratio in the quantile regression algorithm; is the first regression coefficient; is the second regression coefficient; Step 2: Define the time when the short-term load probability forecast of the distribution network is required as the load prediction time t, collect the n distribution network historical load data before the load prediction time t and form the distribution network historical load data set A, as shown in the following formula (1), （2） In formula (2), They are respectively the historical load data of the first distribution network before the load prediction time t to the historical load data of the nth distribution network; Collect the distribution network real load data from the moment before the load prediction moment t to the moment before m and form the distribution network real load data set B, as shown in the following formula (3): （3） In formula (3), are the actual load data of the distribution network from the moment before the load prediction moment t to the m moments before; According to the sliding window method, all data in the distribution network historical load data set A are sliced ​​to form a distribution network historical load data time series sequence from the moment before the load prediction time t to the moment before m+1, and all the distribution network historical load data time series sequences are collected to form a distribution network historical load data time series sequence set X, as shown in the following formula (4): （4） In formula (4), are respectively time series of historical load data of the distribution network from the moment before the load prediction moment t to the moment m+1 before; is the window length of the sliding window; Step 3: Calculate the optimal quantile probability ratio of the initial distribution network short-term load probability forecasting model based on the joint experience pool replay-double delay deep deterministic policy gradient algorithm , the optimal first regression coefficient and the optimal second regression coefficient , the specific steps are as follows: Step 3.1: Establish the quantile probability proportion update model in the initial distribution network short-term load probability forecasting model as shown in the following formula (5): （5） In formula (5), is the updated quantile probability ratio; is the state action value function; x is the real load sequence of the distribution network in the input quantile probability proportional update model; is the quantile probability ratio to be updated in the input quantile probability ratio update model; is a neural network parameter in the initial distribution network short-term load probability forecasting model; is the action space; is the action index value in the action space; is the quantile probability; is the number of actions in the action space H; The regression coefficient update model in the initial distribution network short-term load probability prediction model is established as shown in the following formula (6): （6） In formula (6), is the updated regression coefficient, is the regression coefficient to be updated; Is the gradient calculation symbol; is the empirical loss function; is the learning rate in the initial distribution network short-term load probability forecasting model; is the size of the sampling experience pool; is the correction parameter; is the number of experiences in the predictor buffer; is the empirical sampling probability; is the maximum importance sampling weight; It is the priority of sampling experience; is the degree of sampling experience priority; is the quantile loss; It is the actual load data of the distribution network; is the predicted load probability boundary of the distribution network; The neural network parameter update model in the initial distribution network short-term load probability prediction model is established as shown in the following formula (7): （7） In formula (7), are the updated neural network parameters; are the neural network parameters to be updated; is the update function; is the reward function; Step 3.2: Initialize the quantile probability ratio in the initial distribution network short-term load probability forecasting model and the neural network parameters The value of is used as the quantile probability ratio of the load at the m+1 moments before the load prediction time t and neural network parameters Initialize the first regression coefficient in the initial distribution network short-term load probability forecasting model and the second regression coefficient The value of is used as the first regression coefficient of the initial distribution network short-term load probability forecasting model at the load prediction time t before m+1 time and the second regression coefficient ; Step 3.3: The quantile probability ratio of the initial distribution network short-term load probability forecasting model at the m+1 time before the load prediction time t is calculated. and neural network parameters Enter the quantile probability ratio to update the model, and then The quantile probability ratio update model also outputs the quantile probability ratio of the initial distribution network short-term load probability forecast model at the m moments before the load prediction time t. ; 、 、 and stated Substitute the initial distribution network short-term load probability forecast model into the calculation to obtain the initial load probability upper bound of the distribution network at the m moments before the load prediction time t and the lower bound of the initial load probability ; Step 3.4: 、 、 、 and Substitute the regression update model to calculate the first regression coefficient of the initial distribution network short-term load probability forecast model at the load prediction time t before m times and the second regression coefficient Step 3.5: 、 、 、 、 、 and Substitute into the neural network parameter update model to calculate the neural network parameters of the initial distribution network short-term load probability forecasting model at the m moments before the load prediction time t ; Step 3.6: Based on the principles of steps 3.2 to 3.5, sequentially calculate the quantile probability ratio, the first regression coefficient, the second regression coefficient, and the neural network parameters of the initial distribution network short-term load probability forecasting model from the m-1 moments before the load prediction time t to the previous moment; The quantile probability ratio of the initial distribution network short-term load probability forecasting model at the moment before the load prediction time t is calculated. and neural network parameters Enter the quantile probability ratio to update the model, and then The quantile probability ratio of the initial distribution network short-term load probability forecasting model at the load prediction time t is also output in the quantile probability ratio update model. ; Select the quantile probability ratio of the initial distribution network short-term load probability forecasting model at the load prediction time t As the optimal quantile probability ratio of the initial distribution network short-term load probability forecasting model ; Select the first regression coefficient of the initial distribution network short-term load probability forecasting model at the time before the load prediction time t and the second regression coefficient As the optimal first regression coefficient of the initial distribution network short-term load probability forecasting model and the optimal second regression coefficient ; Step 4: The optimal quantile probability ratio obtained in step 3 , the optimal first regression coefficient and the optimal second regression coefficient Substitute the initial distribution network short-term load probability forecasting model into the final distribution network short-term load probability forecasting model, as shown in the following formula (8): （8） The Input the final distribution network short-term load probability forecast model to calculate the upper limit of the final load probability of the distribution network at the load prediction time t and the lower bound of the final load probability ; The upper bound of the final load probability and the lower bound of the final load probability The interval formed That is, the load probability interval of the distribution network at the load prediction time t.

Images