Method, System, Device and Medium for Predicting the Load of New Users in the Distribution Network by Time Series

Abstract

The present invention relates to the technical field of power grids, and specifically discloses a method, system, device and medium for predicting the time-series load of new users in a distribution network, including collecting the intermittent data of new users during trial operation, using Markov chain and sequential Monte Carlo simulation to extract the probability characteristics of the time-series load state of new users, establishing an interpolation model for the missing time-series load with time and weather as influencing factors, correcting the interpolation model based on a generative adversarial network to obtain the continuous data of new users during trial operation, and using the LS-SVM algorithm, the Kalman filter algorithm or a combination of these two algorithms to predict the time-series load of new users. The present invention can accurately predict the time-series load of new users in the business expansion and installation of distribution networks, which is beneficial to selecting a suitable power access point for new users in the business expansion and installation of distribution networks in practical applications.

Description

Technical Field

[0001] The present invention relates to the technical field of power grids, and particularly to a method, system, device and medium for predicting the time-series load of new users in a distribution network. Background Art

[0002] The load of new users in the new-type distribution network shows diverse and complex variation characteristics, which will have many impacts on the access of newly applied power supply points, "peak shaving and valley filling", the safe and stable operation of the distribution network, and the consumption of renewable energy power generation during the business expansion and installation process of the distribution network.

[0003] The load of newly applied users in the distribution network often lacks a large amount of historical load. Usually, its used equipment only has factory parameters, such as peak value, valley value, rated voltage, rated current, rated power, and the load operation data of the user's power-on test run and intermittent operation for a short time. These data may be incomplete and discontinuous, and problems such as abnormal noise may occur. Therefore, the interpolation and correction of the missing data of the new user load are of great significance for accurately predicting its load.

[0004] At present, the research on the identification and filling of abnormal data of user power consumption load has become the focus and key issue of domestic and foreign research. Power science enthusiasts have proposed a series of solutions to this problem, but the current research does not involve the load of newly applied users in the distribution network, nor does it consider the situation of intermittent operation and short operation time of new user loads. Summary of the Invention

[0005] The purpose of the present invention is to provide a method, system, device and medium for predicting the time-series load of new users in a distribution network, so as to solve the problems that the existing research on the identification and filling of abnormal data of user power consumption load does not involve the load of newly applied users in the distribution network, nor does it consider the situation of intermittent operation and short operation time of new user loads.

[0006] To solve the above technical problems, the present invention provides a method for predicting the time-series load of new users in a distribution network, including:

[0007] Collecting the intermittent data of the new user's power-on test run, and using Markov chain and sequential Monte Carlo simulation to extract the probability characteristics of the time-series load state of the new user from the intermittent data of the power-on test run;

[0008] According to the probability characteristics of the time-series load state of the new user, establishing an interpolation model for the missing time-series load with time and weather as influencing factors;

[0009] Using a generative adversarial network to establish a correction calculation model to correct the interpolation model for the missing time-series load, and obtaining the continuous data of the new user's power-on test run;

[0010] Based on the voltage level, a multi-level three-dimensional topological structure is established with the new user as a node together with other known user nodes. The cross-correlation algorithm is used to calculate the correlation coefficient between the electricity load of the new user node and its upper-level node. The strength of the association between nodes is obtained from the correlation coefficient. Based on the continuous data of the new user's trial operation and the strength of the association between nodes, the LS-SVM algorithm, the Kalman filtering algorithm, or a combination of these two algorithms is used to predict the electricity time-series load of the new user.

[0011] Preferably, the intermittent data of the trial operation is divided into N continuous load states according to the fluctuation period, and these continuous load states form a Markov chain. The transition probability between each state is:

[0012]

[0013] where λ ij is the transition probability between state i and state j, N ij is the number of transitions between state i and state j, T i is the total duration of the entire statistical period. The transition probability matrix containing N continuous load states is:

[0014]

[0015] The extreme value of the state transition probability remains unchanged during multiple transitions:

[0016]

[0017] where are the probabilities of N load states. The durations of N load states are:

[0018]

[0019] where is the average duration of state i, M i is the number of transitions of state i. The probabilities of N load states are obtained based on the transition probability matrix and the durations of N load states, and the sequential Monte Carlo method is used to obtain the actual load state:

[0020]

[0021] where U is a random number subject to a uniform distribution, and S1, S2, …, S N correspond to the actual load states of N respectively. Based on the fact that the duration of the load state follows an exponential distribution, a random number R subject to this distribution is randomly generated to obtain the actual durations of N load states during the trial operation:

[0022]

[0023] Preferably, the fluctuation period is divided into peak period, normal period and off-peak period in a day. The time interval of the peak period is 08:30 - 11:30 and 18:00 - 23:00, the time interval of the normal period is 07:00 - 08:30 and 11:30 - 18:00, and the time interval of the off-peak period is 23:00 - 7:00.

[0024] Preferably, an interpolation model is established with days, hours and weather as influencing factors:

[0025]

[0026] Among them, b 11 , b 12 , b 13 , b 21 , b 22 , b 23 , b3, c k are the coefficients to be regressed, P r is the rated load of the user, is the daily average value of the load, t is the moment, d is the number of days, ω t is the temperature at moment t, ω min is the minimum daily temperature, ω max is the maximum daily temperature, K is the expansion order of the B-spline function, B k (t) is the k-th expansion order of the B-spline function,

[0027] Let the coefficients to be regressed b 11 , b 12 , b 13 , b 21 , b 22 , b 23 , b3, c k be:

[0028]

[0029] The parameter θ is obtained by using the residual estimation model of the optimized quadratic loss function:

[0030]

[0031] Among them, is the new user load data, N d is the number of new user load acquisition data.

[0032] Preferably, a calibration calculation model is established based on the generative adversarial network:

[0033]

[0034] Among them, the function G(x) is a generative function network, the function D(x) is a discriminative function network, E{} is an expectation function, and x is the rated load P of the user. r , P g is the generated load data, is the data in the generated load data P g in,

[0035] There is a loss function:

[0036]

[0037] Among them, ψ ij is the feature vector obtained by the activation function of the th convolutional layer before the th max-pooling layer in the generative network, W ij and H ij are the dimensions of this feature vector, is the weight coefficient, and the generative adversarial network calculation correction model is solved through the loss function.

[0038] Preferably, when there is a weak association between the new user and its upper-level node, the LS-SVM algorithm is used to predict the new user separately. Based on the continuous data of the new user's trial operation, the nonlinear optimization model is established as:

[0039]

[0040] Among them, x is the influencing factor in the prediction process, f(x) is the predicted value, w and b are the model parameters, is the nonlinear mapping of the input parameters, w, b and are learned through the intermittent data of the trial operation during the model training process,

[0041] The nonlinear regression problem that occurs is:

[0042]

[0043] Among them, e i is the error vector element of the n-dimensional fitting, γ is the regularization parameter, and after sorting, we can get:

[0044]

[0045] Among them, ζ = [ζ1, ζ2, …, ζ n T , 1 = [1, 1, …, 1] T , y = [y1, y2, …, y n T is an n-dimensional column vector, I is the identity matrix, is the weight coefficient, G is an n×n matrix H is a kernel function that satisfies the Mercer condition,​​

[0046] The final regression function expression is as follows:

[0047]

[0048] Calculate the predicted value of the new user's power consumption time series load;

[0049] When there is a strong correlation between the new user and its upper-level node, the Kalman filtering algorithm is used for combined prediction. Based on the continuous data of the new user's trial operation and the known power consumption load data of the upper-level node, the discrete state equation X(k) and observation equation Z(k) of the power system are established as follows:

[0050]

[0051] Where X(k) and Z(k) are the system state and measurement value at time k respectively, A is the state transition matrix, u(k) is the system control quantity at time K, B is the system control matrix, H is the measurement system parameter, v(k) is the measurement noise at time k, and w(k) is the process noise at time k.

[0052] Establish a recurrence formula:

[0053]

[0054] Where is the state variable extrapolated forward, is the error covariance extrapolated forward; K k is to calculate the Kalman gain; is the updated optimal estimate; P k is the updated measurement error,

[0055] Calculate the predicted value of the new user's power consumption time series load.

[0056] A power consumption time series load prediction system for new users in a distribution network, including:

[0057] A data collection and analysis module, which is used to collect the intermittent data of the new user's trial operation and extract the probability characteristics of the power consumption time series load state of the new user from the intermittent data of the trial operation by using Markov chain and sequential Monte Carlo simulation;

[0058] An interpolation model establishment and correction module, which is used to establish an interpolation model for the missing power consumption time series load with time and weather as influencing factors according to the probability characteristics of the power consumption time series load state of the new user, establish a correction calculation model by using a generative adversarial network, and correct the interpolation model for the missing power consumption time series load to obtain the continuous data of the new user's trial operation;

[0059] A prediction module for predicting the time-series load of new users using the LS-SVM algorithm, the Kalman filter algorithm, or a combination of these two algorithms.

[0060] A device for predicting the time-series load of new users in a distribution network, including a memory and a processor. The memory is used to store a computer program, and the computer program is used to execute the above method when loaded by the processor.

[0061] A readable storage medium stores a computer program, and the computer program is suitable for executing the above method when loaded by a processor.

[0062] Compared with the prior art, the beneficial effects of the present invention are as follows;

[0063] The method for extracting the time-series load characteristics of newly applied users based on Markov chains and sequential Monte Carlo simulation can accurately obtain the time-series load characteristics of newly applied users. The interpolation method for the missing time-series load of newly applied users considering time and weather can obtain accurate interpolation results. Based on the generative adversarial network, the interpolation model is effectively corrected. Different prediction methods are selected considering the strength of the association between nodes. Using the LS-SVM algorithm to perform nonlinear fitting on the node load data improves the accuracy and calculation speed of the prediction results. Using the Kalman filter algorithm can effectively remove the influence of observation noise and process noise, thereby obtaining a prediction value closer to the true state. The present invention can accurately predict the time-series load of newly applied users in the distribution network business expansion, which is beneficial to selecting a suitable power access point for newly applied users in the distribution network business expansion in practical applications. Description of the Drawings

[0064] Figure 1 It is a three-dimensional topological diagram of the electricity load in the embodiment of the present invention;

[0065] Figure 2 It is an electrical wiring diagram of a distribution network in a certain area in the embodiment of the present invention;

[0066] Figure 3 It is a three-dimensional topological diagram of the electricity load of a distribution network in a certain area in the embodiment of the present invention;

[0067] Figure 4 It is a simulation result diagram of the individual predicted value and the true value of the electricity load of node A in a distribution network in a certain area in the embodiment of the present invention;

[0068] Figure 5 It is a simulation result diagram of the individual predicted value and the true value of the electricity load of node B in a distribution network in a certain area in the embodiment of the present invention;

[0069] Figure 6 It is a simulation result diagram of the individual predicted value and the true value of the electricity load of node C in a distribution network in a certain area in the embodiment of the present invention;

[0070] Figure 7 This is the simulation result graph of the predicted value and the true value of the electricity consumption load at node D of the distribution network in a certain area in the embodiment of the present invention;

[0071] Figure 8 This is the simulation result graph of the value of the electricity consumption load at node C of the distribution network in a certain area in the embodiment of the present invention after Kalman filtering, the observed value, and the true state value;

[0072] Figure 9 This is the simulation result graph of the individual prediction error value at node C of the distribution network in a certain area in the embodiment of the present invention;

[0073] Figure 10 This is the simulation result graph of the combined predicted value and the true value of the electricity consumption load at node B of the distribution network in a certain area in the embodiment of the present invention;

[0074] Figure 11 This is the simulation result graph of the combined prediction error value at node B of the distribution network in a certain area in the embodiment of the present invention;

[0075] Figure 12 This is the simulation result graph of the actual value and the predicted value of the multiple regression at node A of the distribution network in a certain area in the embodiment of the present invention;

[0076] Figure 13 This is the flow chart of the embodiment of the present invention. Specific implementation manners

[0077] The following further describes in detail the implementation manners of the present invention with reference to the drawings and embodiments. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0078] A prediction method for the time - series load of new users in a distribution network in this embodiment is as Figure 13 shown, and includes:

[0079] Collect the intermittent data during the trial operation of new users in the business expansion and installation, as well as the rated power of their electrical equipment, environmental temperature and humidity, and electricity consumption location information, and use the Markov chain and sequential Monte Carlo simulation method to extract their statistical characteristics to obtain the probability characteristics of the time - series load state of new users;

[0080] The fluctuation period in a day is divided into peak period, normal period, and low - valley period. The time interval of the peak period is 08:30 - 11:30 and 18:00 - 23:00, the time interval of the normal period is 07:00 - 08:30 and 11:30 - 18:00, and the time interval of the low - valley period is 23:00 - 7:00. Divide the intermittent data during the trial operation into N continuous load states according to the fluctuation period, and this continuous load state constitutes a Markov chain. The transition probability between each state:

[0081]

[0082] where λ ij is the transition probability between state i and state j, N ij is the number of transitions between state i and state j, T i is the total duration of the entire statistical period, and the transition probability matrix containing N consecutive load states is:

[0083]

[0084] According to the principle of Markov chain, the extreme values of the transition probabilities of N consecutive load states remain unchanged during multiple transitions:

[0085]

[0086] where are the probabilities of N load states, and the duration of N load states is equal to the reciprocal of the sum of state transitions, i.e.:

[0087]

[0088] where is the average duration of state i, M i is the number of transitions of state i. Based on the transition probability matrix and the durations of N load states, the probabilities of N load states are obtained. The sequential Monte Carlo method is used to sample [P1 P2 … P N to obtain the actual load state:

[0089]

[0090] where U is a random number following a uniform distribution, and S1, S2, …, S N correspond to the actual load states of N loads respectively. Based on the fact that the durations of load states follow an exponential distribution, random numbers R following this distribution are randomly generated to obtain the actual durations during the N-load trial operation period:

[0091]

[0092] According to the probability characteristics of the power consumption time-series load states of new users, aiming at the problem that the traditional prediction time scales of years and months are too large and the load prediction based on days is inaccurate, an interpolation model for missing power consumption time-series loads with days, hours, and weather as influencing factors is established:

[0093]

[0094] where b 11 , b 12 , b 13 , b 21 , b 22 , b23 , b3, c k is the coefficient to be regressed, P r is the rated load of the user, is the daily average value of the load, t is the time, the value range is 0 - 24, d is the number of days, ω t is the temperature at time t, ω min is the minimum daily temperature, ω max is the maximum daily temperature, K is the expansion order of the B-spline function, B k (t) is the k-th expansion order of the B-spline function,

[0095] Let the coefficient to be regressed b 11 , b 12 , b 13 , b 21 , b 22 , b 23 , b3, c k is:

[0096]

[0097] Use the residual estimation model of the optimization quadratic loss function to obtain the parameter θ as:

[0098]

[0099] Among them, is the new user load data, N d is the number of new user load acquisition data.

[0100] The generative adversarial network is a type of deep learning network model. Its network structure is mainly composed of an input layer, a convolutional layer, and an output layer, and is mainly divided into two parts: a generative network and an adversarial network. It obtains realistic data by continuously correcting the data. Based on the generative adversarial network, the correction calculation model is established as:

[0101]

[0102] Among them, the function G(x) is the generative function network, the function D(x) is the discriminant function network, E{} is the expectation function, and x is the rated load P of the user r , P g is the generated load data, and x~ is the data in the generated load data P g in,

[0103] There is a loss function:

[0104]

[0105] Among them, ψ ijTo generate the feature vector obtained by the activation function of the th convolutional layer before the th maximum pooling layer in the network, W ij and H ij is the dimension of the feature vector, The loss function is used to solve the generated adversarial network correction calculation model, and the interpolation model of the missing power time series load is corrected to obtain the continuous trial data of new users.

[0106] Topology goes beyond the focus on physical properties such as size, dimensions and shape of objects, and instead uses points and lines to abstractly describe the actual locations of multiple objects and their relationships. The interaction between objects within a specific range is revealed in the form of layers. The energy flow of the power grid presents a multi-level and multi-regional balance state, and its vertical and horizontal connections can be shown through topological structures, such as Figure 1 As shown in the figure, the energy flow topology of the power grid is as follows: for the power grid system constructed in a region, for the voltage level of the system, a balance will be formed between the incoming energy and the outgoing energy of each node. The node load corresponding to the outgoing energy is the incoming energy of the next level of the power grid, and the outgoing energy of the next level of the power grid will form the corresponding node load. This cycle repeats itself, and the entire load system network composed of voltage levels will show a multi-layer, multi-level balanced system.

[0107] Based on the voltage level, the new user is taken as a node and a multi-level three-dimensional topological structure is established together with other known user nodes. The cross-correlation algorithm is used to calculate the correlation coefficient of the power load between the new user node and the node on the previous level. The strength of the association between the nodes is obtained from the correlation coefficient. Based on the continuous data of the new user's trial operation combined with the strength of the association between the nodes, the LS-SVM algorithm, Kalman filtering algorithm or a combination of these two algorithms is used to predict the new user's power load sequence.

[0108] From the perspective of time, when there is a weak correlation between a new user and the node at the previous level, its own regular characteristics are studied, and the LS-SVM algorithm is used to predict the new user separately. The algorithm uses equality constraints instead of traditional inequality constraints, and also takes the square term of the error into account. Since the linear equations are solved, the problem is simplified, which can effectively improve the fitting degree of the node load data in the nonlinear case, and also improve the calculation speed of the algorithm process. Based on the continuous data of the new user trial run, according to the least squares support vector machine method, the nonlinear optimization model is established as follows:

[0109]

[0110] Among them, x is the influencing factor in the prediction process, f(x) is the prediction amount, w and b are model parameters, is the nonlinear mapping of the input parameters, w, b and Obtained through intermittent data learning during the model training process by trial operation.

[0111] The possible non - linear regression problems are as follows:

[0112]

[0113] Among them, e i is the error vector element of the n - dimensional fitting, γ is the regularization parameter, and the purpose is to reduce the fitting error and the model complexity. After arrangement, we can get:

[0114]

[0115] Among them, ζ = [ζ1, ζ2, …, ζ n T , 1 = [1, 1, …, 1] T , y = [y1, y2, …, y n T , which are n - dimensional column vectors, I is the identity matrix, is the weight coefficient, G is an n×n - order matrix, H is the kernel function that satisfies the Mercer condition,

[0116] In summary, the final regression function expression is:

[0117]

[0118] Calculate the predicted value of the electricity consumption time - series load of new users;

[0119] From a vertical perspective in space, when there is a strong correlation between a new user and its upper - level node, consider indirect node prediction. Use the Kalman filtering algorithm for combined prediction through the node allocation factor and the node power factor. By inputting the historical data of the prediction object of input and output, use the optimal algorithm to obtain the system state. Based on the continuous data of the new user's trial operation and the known electricity load data of the upper - level node, establish the discrete state equation X(k) and the observation equation Z(k) of the power system as:

[0120]

[0121] Among them, X(k) and Z(k) are the system state and the measurement value at time k respectively, A is the state transition matrix, u(k) is the system control quantity at time K, B is the system control matrix, H is the measurement system parameter, v(k) is the measurement noise at time k, and w(k) is the process noise at time k.

[0122] Based on the Kalman filter being divided into two parts: time update (prediction) and state update (correction), establish the recurrence formula:

[0123] ​​

[0124] Among them, is the state variable for forward prediction, is the forward prediction error covariance; K k is used to calculate the Kalman gain; is the updated optimal estimate; P k is the updated measurement error;

[0125] The predicted value of the power consumption time series load of the new user is calculated.

[0126] A power distribution network new user power consumption time series load prediction system includes:

[0127] A data acquisition and analysis module, which is used to collect the intermittent data of the new user's trial operation and extract the power consumption time series load state probability characteristics of the new user from the intermittent data of the trial operation by using Markov chain and sequential Monte Carlo simulation;

[0128] An interpolation model establishment and correction module, which is used to establish an interpolation model for the missing power consumption time series load with time and weather as influencing factors according to the power consumption time series load state probability characteristics of the new user, establish a correction calculation model by using a generative adversarial network, and correct the interpolation model for the missing power consumption time series load to obtain the continuous data of the new user's trial operation;

[0129] A prediction module, which is used to predict the power consumption time series load of the new user by using the LS-SVM algorithm, the Kalman filter algorithm or a combination of these two algorithms.

[0130] A power distribution network new user power consumption time series load prediction device includes a memory and a processor. The memory is used to store a computer program, and the computer program is used to execute the above method when loaded by the processor.

[0131] A readable storage medium stores a computer program, and the computer program is suitable for executing the above method when loaded by the processor.

[0132] The present invention and its effects are specifically described below through an example.

[0133] Suppose there is a network structure diagram as Figure 2 shown in a certain area's power distribution network. This structure is composed of four loads, namely A, B, C, and D, through the network space. The voltage level of load A is 220 kV, and the voltage levels of loads B, C, and D are 110 kV. Among them, node B is a new user for power distribution network business expansion and installation, and node B has performed intermittent data interpolation according to the method described above. The interpolation results are shown in Table 1.

[0134] Table 1 Time series interpolation results of the newly installed user B

[0135]

[0136] On this basis, the electricity load of new users of Node B is predicted. At the same time, in order to verify the effectiveness of the prediction method proposed in this embodiment, the electricity loads of Nodes A, C, and D are predicted and compared.

[0137] First, establish a three-dimensional topological diagram of the electricity load between nodes as shown in Figure 3 and calculate the correlation between the four nodes using the cross-correlation algorithm. The results are shown in Table 2.

[0138] Table 2 Correlation coefficients between four loads

[0139]

[0140] As can be seen from Table 2, the correlations between Nodes C and D and Node A are strong. The prediction method based on the Kalman filtering method in the vertical direction can be used. While the correlation between Node B and Node A is between strong and weak, the node combination prediction method can be used horizontally.

[0141] Next, first study the electricity load characteristics of Nodes A, B, C, and D. First, use the LS-SVM algorithm to make a separate self-prediction for each load, as shown in Figures 4-7 . In the figure, "△" represents the predicted value, and "○" represents the true value. Calculate the prediction error according to the predicted value, as shown in Table 3.

[0142] Table 3 Average relative errors of predictions for Nodes A, B, C, and D

[0143]

[0144] By comparing the average relative errors of the predicted values of the electricity loads of Nodes A, B, C, and D, it is found that the average relative error of Node C is relatively large.

[0145] Then, the Kalman filtering algorithm is applied to Node C to reduce the large error in the separate prediction method using the LS-SVM algorithm. According to the optimal estimated predicted value and the optimal covariance matrix predicted by the upper-level Node A, estimate the prior estimated value and the prior estimated covariance matrix of Node C. During the prediction process of Node C, there may be a process error due to the prediction interference of Node A and a measurement error during its own prediction process. According to the Kalman recursion formula, first calculate the Kalman gain, and then combine the prior estimated value and the observation equation to obtain the corrected optimal estimate and update the optimal estimated covariance matrix. Perform cyclic prediction and correction in this way to obtain the optimal estimated predicted value.

[0146] During the prediction process, the initial state is determined, and state prediction and observation prediction are performed on the target. Combining the recurrence formula in the Kalman augmentation filtering process, the simulation prediction result is obtained as Figure 8 shown. The average relative error is shown in Table 4. It can be seen that the error value of the prediction result of Node C is smaller than that of the prediction using the LS-SVM algorithm alone.

[0147] Table 4 Average Relative Error of Node C

[0148]

[0149] For the nodes with the association between strong and weak, the weighted calculation method is used. In this embodiment, the weights of both the LS-SVM algorithm and the Kalman filtering algorithm are 0.5. Different weight values can also be assigned according to the preference for different algorithms. For the node to be predicted, Node B, the restraint effect of the adjacent Node C is utilized, and the calculated value of Node B under the prediction using the LS-SVM algorithm alone and the numerical value of Node C based on the Kalman filtering algorithm are combined for prediction. The prediction result is as Figures 9-11 shown. The average relative error is shown in Table 5.

[0150] Table 5 Average Relative Error of Node B

[0151]

[0152] It can be seen that the predicted value using the combination of the LS-SVM algorithm and the Kalman filtering algorithm is closer to the actual value, the fluctuation of the error value is relatively small, the overall result is close to the actual value, and it can meet the effective prediction of the power system.

[0153] In addition, in the entire energy flow topology, multiple regression is performed on the four nodes A, B, C, and D. The data of nodes B, C, and D are input for regression prediction and compared with the actual value of node A. The data selects the data of 1 day at the same moment for prediction and processing. The simulation result is as Figure 12 shown. The average relative error is shown in Table 6.

[0154] Table 6 Average Relative Error of Node A

[0155]

[0156] Through the above prediction process, the calculation data of the three load nodes A, B, and C are recorded, as shown in Table 7.

[0157] Table 7 Calculation Data of Nodes A, B, and C

[0158]

[0159] From the comparison of the calculation results, it can be seen that the electricity load forecasting method proposed in this embodiment is closer to the actual situation, and the forecasting accuracy is more remarkable.

[0160] The method for extracting the time-series load characteristics of newly installed users based on Markov chain and sequential Monte Carlo simulation can accurately obtain the time-series load characteristics of newly installed users. The interpolation method for the missing time-series load of newly installed users considering time and weather can obtain accurate interpolation results. Based on the generative adversarial network, the interpolation model is effectively corrected. Different forecasting methods are selected according to the strength of the association between nodes. The LS-SVM algorithm is used to perform nonlinear fitting on the node load data, improving the accuracy and calculation speed of the forecasting results. The Kalman filter algorithm can effectively remove the influence of observation noise and process noise, so as to obtain a forecasting value closer to the true state. The present invention can accurately forecast the time-series load of newly installed users in the distribution network for business expansion, which is beneficial to selecting a suitable power access point for newly installed users in the distribution network for business expansion in practical applications.

[0161] The embodiments of the present invention are given for purposes of illustration and description, and are not exhaustive or limit the invention to the disclosed form. Many modifications and variations are obvious to those of ordinary skill in the art. The embodiments are selected and described to better illustrate the principles and practical applications of the present invention, and to enable those of ordinary skill in the art to understand the present invention and design various embodiments with various modifications suitable for specific purposes.

Claims

1. A prediction method for the electricity consumption time-series load of new users in a distribution network, characterized in that, Including: Collect the intermittent data of new users during the trial operation, and use Markov chain and sequential Monte Carlo simulation to extract the probability characteristics of the power consumption time-series load status of new users from the intermittent data of the trial operation; According to the probability characteristics of the power consumption time-series load status of new users, establish an interpolation model for the missing power consumption time-series load with time and weather as influencing factors; Use a generative adversarial network to establish a calibration calculation model to calibrate the interpolation model for the missing power consumption time-series load, and obtain the continuous data of new users during the trial operation; Based on the voltage level, establish a multi-level three-dimensional topological structure with the new user as a node and other known user nodes. Use the cross-correlation algorithm to calculate the correlation coefficient between the power consumption load of the new user node and its upper-level node. The strength of the association between nodes is obtained from the correlation coefficient. Based on the continuous data of the new user during the trial operation and the strength of the association between nodes, use the LS-SVM algorithm, the Kalman filter algorithm, or a combination of these two algorithms to predict the power consumption time-series load of the new user.

2. The prediction method for the power consumption time series load of new users in the distribution network according to claim 1, characterized in that, Collect the intermittent data of new users during the trial operation, and use Markov chain and sequential Monte Carlo simulation to extract the probability characteristics of the power consumption time-series load status of new users from the intermittent data of the trial operation, including: Divide the trial operation intermittent data into continuous load states according to the fluctuation period, and these continuous load states form a Markov chain. The transition probabilities between each state are: ； wherein, is the and transition probability between states, is the and number of transitions between states, is the total duration of the entire statistical period, and the transition probability matrix containing consecutive load states is: ； The extreme values of the state transition probability remain unchanged during multiple transitions: ； Among them, is the probability of a load state, and the duration of a load state is: ； Among them, is the average duration of the state , is the number of transitions of the state . It is obtained according to the transition probability matrix and the duration of load states to get load state probabilities. The actual load state is obtained using the sequential Monte Carlo method: ； Among them, is a random number subject to a uniform distribution, corresponding to the actual states of N loads respectively, and according to the duration of the load states subject to an exponential distribution, random numbers subject to this distribution are randomly generated , and the actual durations during the trial operation of the loads are obtained: 。 3. The prediction method for the power consumption time-series load of new users in the distribution network according to claim 2, wherein The fluctuation period is divided into peak period, normal period, and valley period in a day. The peak period time interval is 08:30~11:30, 18:00~23:00, the normal period time interval is 07:00~08:30, 11:30~18:00, and the valley period time interval is 23:00~7:

00.

4. The prediction method for the power consumption time-series load of new users in a distribution network according to claim 2, characterized in that Establish an interpolation model with days, hours, and weather as influencing factors: ； Among them, is the coefficient to be regressed, is the rated load of the user, is the daily average value of the load, is the time, is the number of days, is the time temperature of, is the minimum daily temperature, is the maximum daily temperature, is the expansion order of the spline function, is the th expansion order of the spline function, Let the regression coefficient to be estimated be ； Obtain parameters using a residual estimation model with an optimized quadratic loss function Namely: ； Among them, is the load data of new users, is the number of load acquisition data of new users.

5. The prediction method for the power consumption time-series load of new users in the distribution network according to claim 3, characterized in that Establish a calibration calculation model based on a generative adversarial network: ； Among them, the function is the generation function network, the function is the discriminant function network, is the expected function, is the rated load of the user , is the generated load data, is the generated load data in it, There is a loss function: ； Among them, is the feature vector obtained by the activation function of the i-th convolutional layer before the i-th max-pooling layer in the generation network, and is the dimension of this feature vector, is the weight coefficient, and the correction calculation model of the generative adversarial network is solved through the loss function.

6. The prediction method for the power consumption time-series load of new users in a distribution network according to claim 5, wherein, When there is a weak association between the new user and its upper-level node, use the LS-SVM algorithm to predict the new user alone. Based on the continuous data of the new user during the trial operation, establish a non-linear optimization model as: ； Among them, is an influencing factor during the prediction process, is the predicted quantity, and are model parameters, is the non-linear mapping of the input parameters, 、 and are obtained by learning intermittent data through trial runs during the model training process, The non-linear regression problem that occurs is: ； Among them, is the error vector element resulting from the dimensional fitting. After arrangement, we can obtain: ； Among them, is a column vector of dimension is the identity matrix, is the weight coefficient, is a matrix of order , is a kernel function satisfying the Mercer condition The final regression function expression obtained is: ； Calculate the predicted value of the power consumption time-series load of the new user; When there is a strong correlation between a new user and its upper-level node, the Kalman filter algorithm is used for combined prediction. Based on the continuous data of the new user's trial operation and the known power consumption load data of the upper-level node, a discrete state equation of the power system is established. and the observation equation are as follows: ； wherein, and are the system state and measurement value at time k respectively, is the state transition matrix, is the system control quantity at time K, is the system control matrix, is the measurement system parameter, is the measurement noise at time k, is the process noise at time k, Establish a recurrence formula: ； Among them, is the state variable for forward prediction, is the error covariance for forward prediction; is to calculate the Kalman gain; is the updated optimal estimate; is the updated measurement error, Calculate the predicted value of the power consumption time-series load of the new user.

7. A power consumption time-series load forecasting system for new users in a distribution network, characterized in that, Including: A data collection and analysis module, which is used to collect the intermittent data of new users during the trial operation and use Markov chain and sequential Monte Carlo simulation to extract the probability characteristics of the power consumption time-series load status of new users from the intermittent data of the trial operation; An interpolation model establishment and calibration module, which is used to establish an interpolation model for the missing power consumption time-series load with time and weather as influencing factors according to the probability characteristics of the power consumption time-series load status of new users, use a generative adversarial network to establish a calibration calculation model, and calibrate the interpolation model for the missing power consumption time-series load to obtain the continuous data of new users during the trial operation; A prediction module, which is used to use the LS-SVM algorithm, the Kalman filter algorithm, or a combination of these two algorithms to predict the power consumption time-series load of the new user.

8. A device for predicting the power consumption time-series load of new users in a distribution network, characterized in that, Including a memory and a processor. The memory is used to store a computer program, and the computer program is used to execute the method described in any one of claims 1-6 when loaded by the processor.

9. A readable storage medium, characterized in that, A computer program is stored in the storage medium, and the computer program is adapted to execute the method according to any one of claims 1-6 when loaded by a processor.

Images