An Adaptive Integrated Load Forecasting Method for Distribution Networks Oriented to Edge Computing

Abstract

An adaptive integrated load forecasting method for a distribution network oriented to edge computing, comprising: inputting relevant information of the area under the jurisdiction of the edge computing device of the distribution network; respectively performing standardization processing on the load historical data set and the environmental temperature historical data set in the initial training set, and obtaining a training set; performing dimensionality reduction on the reservoir echo state variables; obtaining the integrated echo state network prediction model for the first day; obtaining the load prediction result for the nth day; obtaining the integrated echo state network prediction model for the (n + 1)th day; determining whether the long-term prediction days identifier is less than or equal to the long-term prediction days identifier; obtaining new model weight coefficients. The present invention not only gives play to the generalization advantage of the integrated learning model, but also avoids the redundant calculation of the integrated learning model; further reduces the computing and storage resources during online update; solves the adaptive problem of the edge-side load prediction model, covers the entire prediction cycle to minimize the model complexity.

Description

Technical Field

[0001] The present invention relates to a method for predicting the daily load of a distribution network. In particular, it relates to an adaptive integrated load prediction method for a distribution network oriented to edge computing. Background Art

[0002] Due to the characteristics of convenient conversion with other forms of energy, easy mass production, centralized management, and long-distance transmission, electric energy has become one of the most widely used energy forms in human life. However, electric energy cannot be stored in large quantities, and the power generation, transmission, transformation, distribution, and consumption of the power system are almost carried out simultaneously. Excessive power generation will cause energy waste, while insufficient power generation may lead to local power outages and even the collapse of the entire system. Therefore, accurate load prediction is crucial for formulating power generation strategies and generating control instructions for various voltage regulation and frequency modulation devices, and is the prerequisite for ensuring that the power system is always in dynamic balance. Among them, the daily load prediction is a basic tool for generating control strategies for distributed power sources, determining the tap ratio of on-load tap changers, thereby improving the new energy consumption rate and ensuring the safety and stability of the system.

[0003] With the rapid development of smart grids, the massive historical data accumulated by distribution networks provides data support for accurate daily load prediction. However, in the traditional cloud computing architecture, the data transmission process from the access point to the cloud computing center occupies most of the latency. Considering that the data needs to pass through multiple levels of routing in the backbone network, this latency is almost inevitable. At the same time, as the data aggregation hub, the cloud bears a huge computing pressure. Therefore, it is an inevitable requirement for large-scale real-time computing to move computing resources from the cloud center to network edge devices close to users. Establishing an intelligent distribution network with edge computing as the core not only completely avoids data transmission latency and loss in the wide area network, but also improves the privacy security level and access efficiency of data, making in-situ load prediction on edge devices more real-time and accurate, greatly improving the flexibility of service deployment and management, and empowering the power system more precisely.

[0004] Different from system-level load prediction, in-situ load prediction on the edge side faces the following new difficulties and challenges:

[0005] 1) The prediction model for edge computing not only pursues high accuracy, but also must minimize the computational overhead and storage resources of the prediction model in order to be possibly deployed on edge terminal devices for implementation;

[0006] 2) The objects of edge-side load prediction range from the load of a single user or an electric vehicle charging pile to the total load of a substation area, and the load size is generally in the order of several hundred kilowatts. Therefore, the randomness of edge-side load is stronger and the change is more frequent, which requires the prediction model to have good generalization;

[0007] 3) The hyperparameter optimization of the prediction model generally uses heuristic algorithms, which require frequent retraining of the model, resulting in high computational overhead and wasting computational resources.

[0008] 4) How to utilize the streaming data continuously collected on the edge side, give full play to the value of new data, adaptively update the key parameters of the prediction model, and avoid retraining of the prediction model.

[0009] For the artificial intelligence load prediction model on the edge side, existing research mainly focuses on model compression of lightweight neural networks, trying to balance model complexity and prediction performance as much as possible. Currently, the widely studied lightweight neural networks mainly include MobileNets, ShuffleNets, separable convolutional networks, etc. The model compression techniques mainly include pruning the weights of lightweight networks and dimensionality reduction of hidden layer variables. However, the above lightweight networks are mostly trained using the gradient descent algorithm, with many updated parameters during weight iteration and numerous hyperparameters. The entire optimization process has a large computational amount and may have the problem of gradient disappearance. The Echo State Network (ESN), as a new type of Recurrent Neural Network (RNN), not only retains the advantages of RNN in processing time series but also does not require the use of the backpropagation algorithm for weight iteration. It only needs to use the least squares estimation to output the weight matrix, greatly improving the training efficiency. At the same time, ESN has few hyperparameters, and the main hyperparameters are only the leakage rate, reservoir size, and regularization parameter, with a small computational amount for hyperparameter optimization.

[0010] Due to the random initialization of the initial parameters of ESN and the one-time training mechanism, the prediction of unknown data by a single ESN prediction model often shows unstable situations. Ensemble learning, as a meta-algorithm, is not a separate machine learning algorithm but completes the learning task by constructing and combining multiple machine learners (base learners). Therefore, when the load changes frequently, the ensemble prediction model can often obtain higher prediction accuracy and generalization than a single prediction model. However, the ensemble learning algorithm needs to integrate numerous base learners, and the computational resources consumption and storage resources in practical applications are also huge (depending on the ensemble scale). Therefore, to adapt to the hardware resource constraints of edge computing devices, it is necessary to sparsify numerous base learners. At the same time, affected by the external environment, the distribution of load data and influencing factors changes over time. Transfer learning, as an auxiliary framework, provides an important breakthrough for solving the problem of data drift. In terms of maintaining the adaptability of the prediction model, the Kalman filter algorithm, as a data assimilation method, can incorporate new measurement data into the original prediction model to achieve dynamic prediction of the load. Moreover, the Kalman filter uses the linear system state equation to optimally estimate the system state through the system input-output observation data, which is completely compatible with the structure of ESN. Summary of the Invention

[0011] The technical problem to be solved by the present invention is to provide an adaptive integrated load forecasting method for a distribution network oriented to edge computing that can achieve a reduction in the final integration scale in order to overcome the deficiencies of the existing technology.

[0012] The technical solution adopted by the present invention is: an adaptive integrated load forecasting method for a distribution network oriented to edge computing, including the following steps:

[0013] 1) Input the load historical data set E and the environmental temperature historical data set T of the area under the jurisdiction of the distribution network edge computing device as the initial training set Ω; set the number of initial training samples n0, the number of training samples Γ, and the candidate set of leakage rates Candidate set of reservoir parameters Candidate set of regularization parameters Integration scale N0 = N1 × N2 × N3, low-dimensional dimension D of the echo state variable L , number of days M in the source domain d , number of days N in the target domain d , set {P} of covariance matrices of the initial posterior estimation error, set {Q} of initial process noise covariance matrices, initial measurement noise covariance vector R M , number k of training samples for calculating the correction factor, long-term prediction days threshold T L ;

[0014] Among them, N1, N2, and N3 are the total numbers of the candidate sets where the leakage rate, reservoir parameters, and regularization parameters are located respectively; is the N1th leakage rate in the leakage rate candidate set, is the N2th reservoir parameter in the reservoir parameter candidate set, is the N3th regularization parameter in the regularization parameter candidate set, and the integration scale N0 is the total number of basic echo state networks;

[0015] 2) Perform min-max normalization processing on the load historical data set E and the environmental temperature historical data set T in the initial training set Ω respectively, and perform training sample partitioning to obtain the training set Ω′;

[0016] 3) Based on the limitation of the computing power of the edge computing device, select the echo state network in the lightweight neural network as the core prediction algorithm, and introduce sparse coding to reduce the dimension of the reservoir echo state variable. Based on the training set Ω′, train N0 basic echo state networks {G l (x), l = 1, 2,..., N0} respectively; among them, G l (x) is the lth basic echo state network;

[0017] 4) Set the current prediction day number \(n = 1\) and the long-term prediction day number identifier \(c = 1\). To reduce the storage resource requirements and integration scale of the edge computing device, for the \(N_0\) basic echo state networks \(\{G l (x), l = 1, 2, \ldots, N_0\}\) obtained by training, use the Tradaboost algorithm of sparse adaptive boosting to automatically delete non-critical basic echo state networks, and for the remaining basic echo state networks \(\{G 1,z (x), z = 1, 2, \ldots, N sp \}\) and the corresponding model weight coefficients \(\{\alpha z , z = 1, 2, \ldots, N sp \}, integrate to obtain the integrated echo state network prediction model \(G 1 (x)\) for the first day;

[0018] Among them, \(G 1,z (x)\) is the \(z\)-th basic echo state network of the integrated echo state network prediction model \(G 1 (x)\) for the first day, and \(N sp \) is the number of remaining basic echo state networks after automatic screening;

[0019] 5) Before the start of the current prediction day, that is, before the \(n\)-th day, use the integrated echo state network prediction model \(G n (x)\) for day-ahead prediction, input the actual load value of the day before the prediction day, the actual load values of the 7 days before the prediction day, and the ambient temperature data of the weather forecast for the prediction day, and obtain the load prediction result for the \(n\)-th day

[0020] 6) After the \(n\)-th day ends, record the actual load data for the \(n\)-th day To avoid frequent retraining of the model to further improve the prediction speed, use the Kalman filter algorithm with modified covariance to update the parameters of the integrated echo state network prediction model \(G n (x)\) for the \(n\)-th day to obtain the integrated echo state network prediction model \(G n+1 (x)\) for the \((n + 1)\)-th day;

[0021] 7) If \(c \leq T L , then let \(n = n + 1\), \(c = c + 1\), and return to step 5); otherwise, proceed to step 8);

[0022] 8) Based on the prediction errors of each basic echo state network \(\{G n (x), z = 1, 2, \ldots, N n,z \}\) of \(G sp \) from the \((n - T L + 1)\)-th day to the \(n\)-th day, use the Adaboost algorithm to n each basic echo state network \(\{Gn,z (x), z = 1, 2, …, N sp} of the model weight coefficients {α z} are recalculated to obtain new model weight coefficients {α′ z}, let n = n + 1, c = 1, and return to step 5); where G n,z (x) represents the z-th basic echo state network in the integrated echo state network prediction model G n (x) on the n-th day.

[0023] An adaptive integrated load forecasting method for a distribution network oriented to edge computing according to the present invention is based on solving problems such as strong load randomness, obvious seasonality, and self - adaptability of the prediction model in local load forecasting on the edge side, and at the same time, as much as possible reducing the complexity of the prediction model and the occupation of storage resources. In the initial training stage of the model, the TrAdaBoost algorithm is used as the main integration framework, and the ESN is used as the basic predictor; during the integration process, by deleting ESN predictors with low prediction performance and correspondingly improving ESN predictors with high prediction performance, the final integration scale is reduced, reducing the computational overhead and storage resources of subsequent predictions. In the daily rolling prediction stage, combined with the actual load data of each day, the Kalman filter algorithm with modified covariance is used to update the parameters of each basic ESN predictor; when the threshold for long - term prediction is reached, combined with the prediction performance of each basic predictor in the previous period, the Adaboost algorithm is used to redistribute the weights of each basic ESN predictor, solving the self - adaptability problem of the edge - side load prediction model, and providing a basis for refined control optimization such as generating control strategies for distributed power sources in the distribution network and determining the transformation ratio of on - load tap - changing transformers. The main advantages of the present invention are as follows:

[0024] (1) In the initial training stage of the model, the ESN is used as the basic predictor, and the TrAdaBoost algorithm is used as the main integration framework; during the integration process, by deleting ESN predictors with low prediction performance and correspondingly improving ESN predictors with high prediction performance, the final integration scale is reduced, reducing the computational overhead and storage resources of subsequent predictions. It not only gives play to the generalization advantage of the integrated learning model but also avoids the redundant calculation of the integrated learning model;

[0025] (2) In the daily rolling prediction stage, to give full play to the value of newly collected streaming data, combined with the actual load data of each day, the Kalman filter algorithm with modified covariance is used to update the parameters of each basic ESN predictor; the dimensionality of the hidden - layer variables is reduced through an auto - encoder, further reducing the computational and storage resources during online update;

[0026] (3) When the threshold of long-term prediction is reached, the weights of each basic ESN predictor are redistributed by using the AdaBoost algorithm in combination with the prediction performance of each basic predictor in the previous period, so as to solve the adaptive problem of the edge-side load prediction model and cover the entire prediction cycle to minimize the model complexity;

[0027] (4) In terms of hardware implementation, the prediction program can be deployed on an actual edge computing device with only 1GB of memory and an 800MHz CPU, while many deep learning models have high complexity and are difficult to be deployed on actual edge computing devices. Description of the Drawings

[0028] Figure 1 is a flowchart of a method for adaptive integrated load prediction of a distribution network for edge computing according to the present invention;

[0029] Figure 2 is a comparison chart of prediction results of different methods in Area 1 of Example 1;

[0030] Figure 3 is a comparison chart of prediction results of different methods in Area 20 of Example 2. Detailed Embodiments

[0031] The following will make a detailed description of a method for adaptive integrated load prediction of a distribution network for edge computing according to the present invention in combination with the embodiments and the drawings.

[0032] As shown in the Figure 1 drawings, a method for adaptive integrated load prediction of a distribution network for edge computing according to the present invention includes the following steps:

[0033] 1) Input the historical load data set E and the historical environmental temperature data set T of the area under the jurisdiction of the edge computing device of the distribution network as the initial training set Ω; set the number of initial training samples n0, the number of training samples Γ of the Echo State Network (ESN), the candidate set of leakage rates the candidate set of reservoir parameters the candidate set of regularization parameters the integration scale N0 = N1 × N2 × N3, the low-dimensional dimension D of the echo state variable L , the number of days M in the source domain d , the number of days N in the target domain d , the set {P} of the covariance matrices of the initial posterior estimation errors, the set {Q} of the initial process noise covariance matrices, the initial measurement noise covariance vector R M , the number k of training samples for calculating the correction factor, the long-term prediction days threshold T L ;

[0034] Wherein, N1, N2, and N3 are respectively the total numbers of the candidate sets where the leakage rate, the reservoir parameters, and the regularization parameters are located; is the N1-th leakage rate in the leakage rate candidate set, is the N2-th reservoir parameter in the reservoir parameter candidate set, is the N3-th regularization parameter in the regularization parameter candidate set, and the integration scale N0 is the total number of the basic echo state networks;

[0035] 2) Perform min-max normalization processing on the load historical data set E and the environmental temperature historical data set T in the initial training set Ω respectively, and perform training sample partitioning to obtain the training set Ω′;

[0036] The described training sample partitioning to obtain the training set Ω′ is specifically as follows: Based on the normalized initial training set Ω, perform training sample partitioning. The input data of the i-th training sample is: the actual load value 1 day before the i-th prediction day, the actual load values of the previous 7 days, and the environmental temperature data of the weather forecast on the i-th prediction day, and merge the above input data into a column vector u(i), and the dimension is denoted as M×1; the label of the i-th training sample is the actual load value y(i) on the i-th prediction day in the training set Ω′, and the dimension of the label of the training sample is denoted as N×1; wherein, N is the total number of time periods for day-ahead prediction.

[0037] The min-max normalization calculation formula is as follows:

[0038]

[0039] In the formula, the data type U∈{E,T}, U represents the load historical data set E and the environmental temperature historical data set T corresponding to the data type U, and U i represents the i-th data in U, represents the value after normalization, and [a, b] represents the numerical interval after normalization, and here it takes [0, 1];

[0040] 3) Due to the limitation of the computing power of the edge computing device, select the echo state network in the lightweight neural network as the core prediction algorithm, and introduce sparse coding to reduce the dimension of the reservoir echo state variables. Based on the training set Ω′, train N0 basic echo state networks {G l (x), l = 1, 2, …, N0} respectively; wherein, G l (x) is the l-th basic echo state network;

[0041] The described training of N0 basic echo state networks {G l (x), l = 1, 2, …, N0} respectively based on the training set Ω′ is specifically as follows:

[0042] (3.1) Permute the leakage rate candidate set the candidate set of reservoir parameters the candidate set of regularization parameters and use them as the hyperparameters of N0 basic echo state networks;

[0043] (3.2) Set the hyperparameter group of the l-th basic echo state network G l (x) as {r l , L l , C l}. Initialize the reservoir echo state variable x l (0) = 0, and randomly generate the input weight matrix the internal connection matrix of the reservoir the sparse coding random weight matrix A l , the random bias matrix B l ; where r l , L l , C l represent the leakage rate, reservoir parameters, and regularization parameters of G l (x) respectively, x l is of dimension L l × 1, is of dimension L l × M, is of dimension L l × L l , A l is of dimension D L × (M + L l ), B l is of dimension D L × Γ, M is the dimension of the input column vector u of a single training sample, and Γ is the number of training samples;

[0044] Among them, the input weight matrix the random sparse matrix the sparse coding random weight matrix A l , the random bias matrix B l are each generated by random numbers uniformly distributed on [-1, 1]; the random sparse matrix and the internal connection matrix of the reservoir have the same dimension, is obtained as follows:

[0045]

[0046] In the formula, is the scaling factor. The closer it is to 1, the stronger the echo characteristic of the ESN. Here |λ| max represents The maximum value of the modulus of all eigenvalues;

[0047] (3.3) Based on the first n0 initial training samples in the training set Ω′, recursively calculate the reservoir echo state variables of the l-th basic echo state network G l (x), and the recurrence formula is as follows:

[0048]

[0049] where u(i) is the input vector of the i-th training sample in the training set Ω′, and x l (i - 1), x l (i) are the reservoir echo state variables of the (i - 1)-th and i-th training samples in the training set Ω′ respectively, and tanh is the hyperbolic tangent function;

[0050] (3.4) Form:

[0051] (M + L l )×1-dimensional column vector s l (i) = [u(i); x l (i)],

[0052] (M + L l )×Γ-dimensional matrix S l = [s l (n0 + 1), s l (n0 + 2), …, s l (n0 + Γ)],

[0053] N×Γ-dimensional matrix Y l = [y(n0 + 1), y(n0 + 2), …, y(n0 + Γ)], where y(n0 + 1) is the actual load value of the (n0 + 1)-th training sample in the training set Ω′;

[0054] Calculate the low-dimensional feature matrix H l and the sparse coding dimensionality reduction matrix the output weight matrix Complete the training of the l-th basic echo state network G l (x):

[0055] H l = A l S l + B l (4)

[0056]

[0057]

[0058] where I is the identity matrix, and the dimension of H l is DL × Γ, with dimension D L × (M + L l ), with dimension N × D L ;

[0059] (3.5) Repeat steps (3.2) to (3.4) until the training of N0 basic echo state networks is completed.

[0060] 4) Set the current prediction day number n = 1 and the long - term prediction day number flag c = 1. To reduce the storage resource requirements and integration scale of the edge computing device, for the N0 basic echo state networks {G l (x), l = 1, 2, …, N0} obtained by training, use the Tradaboost algorithm with sparse adaptive boosting to automatically delete non - critical basic echo state networks. The remaining basic echo state networks {G 1,z (x), z = 1, 2, …, N sp} and the corresponding model weight coefficients {α z , z = 1, 2, …, N sp} are integrated to obtain the integrated echo state network prediction model G 1 (x) for the first day;

[0061] Among them, G 1,z (x) is the z - th basic echo state network of the integrated echo state network prediction model G 1 (x) for the first day, and N sp is the number of remaining basic echo state networks after automatic screening;

[0062] The method of using the Tradaboost algorithm with sparse adaptive boosting to automatically delete non - critical basic echo state networks, and integrating the remaining basic echo state networks {G 1,z (x), z = 1, 2, …, N sp} and the corresponding model weight coefficients {α z , z = 1, 2, …, Nsp to obtain the integrated echo state network prediction model G1x for the first day specifically includes:

[0063] (4.1) Divide the Γ training samples used for training into a source domain data set composed of the first M d training samples and a target domain data set composed of the remaining N d training samples in chronological order; Let the source domain weight parameter The current iteration number l = 1, and calculate the initial weights of each training sample in the training set Ω′

[0064]

[0065] (4.2) For the training set Ω′ with the sample weight distribution D l-1 calculate the maximum regression error E of the training samples of the l-th basic echo state network G l (x) on the training set Ω′ l :

[0066] E l = max (∑|y(i) - G l (u(i))|), i = 1, 2, …, M d + N d (8)

[0067] where G l (u(i)) represents the predicted value of the l-th basic echo state network G l (x) for the i-th training sample in the training set Ω′, y(i) is the actual value of the i-th training sample in the training set Ω′, and both the dimensions of G l (u(i)) and y(i) are N×1. ∑|y(i) - Gl(ui)| represents the sum of the absolute values of all elements of the vector yi - Gl(ui);

[0068] (4.3) Calculate the relative regression error e of each training sample of the l-th basic echo state network G l (x) on the training set Ω′ l,i :

[0069]

[0070] (4.4) Calculate the overall regression error e of the l-th basic echo state network G l (x) on the training set Ω′ and the target domain weight parameter β l and l :

[0071]

[0072]

[0073] where represents taking the relatively smaller value between and 0.5,

[0074] (4.5) Calculate the model weight coefficient α l of the l-th basic echo state network G l (x) and the cumulative model weight coefficient R:

[0075]

[0076]

[0077] where α v represents the model weight coefficient of the v-th basic echo state network G v (x) selected at the v-th iteration;

[0078] (4.6) If l = 1, then set l = l + 1 and return to step (4.2); otherwise, find the basic echo state networks at the h1-th and h2-th iterations for which the model weight coefficients need to be adjusted according to the following formula, denoted as and Then go to step (4.7):

[0079]

[0080]

[0081] where u(i) represents the input vector of the i-th training sample in the training set Ω′, and G h (u(i)) represents the predicted value of the i-th training sample u(i) in the training set Ω′ by the h-th basic echo state network G h (x) selected at the h-th iteration, and G v (u(i)) represents the predicted value of the i-th training sample u(i) in the training set Ω′ by the v-th basic echo state network G v (x) selected at the v-th iteration; represents the Hadamard product, that is, the corresponding elements of the matrices are multiplied;

[0082] (4.7) Calculate the optimal model weight coefficient transfer step p * and perform the corresponding adjustment of the model weight coefficients of the basic echo state network:

[0083]

[0084]

[0085] where represents the predicted value of the i-th training sample in the training set Ω′, represents the predicted value of the i-th training sample in the training set Ω′, represents the model weight coefficient of represents the model weight coefficient of

[0086] (4.8) If Then, the cumulative model weight coefficient R is adjusted using the following formula (16), and the weight coefficients of the basic echo state network models with model weight coefficients less than 0 are set to 0, and the basic echo state network is deleted:

[0087]

[0088] (4.9) Update the training sample weights of the source domain and the target domain:

[0089]

[0090] (4.10) If l < N0, then let l = l + 1 and return to step (4.2); otherwise, calculate the integrated echo state network prediction model G 1 (x) for the first day:

[0091]

[0092] 5) Before the start of the current prediction day, that is, before the nth day, use the integrated echo state network prediction model G n (x) to perform day-ahead prediction. Input the actual load value of the day before the prediction day, the actual load values of the seven days before the prediction day, and the ambient temperature data of the weather forecast for the prediction day to obtain the load prediction result for the nth day including:

[0093] (5.1) Set the integrated echo state network prediction model G n (x) to be integrated by N sp basic echo state networks {G n,z (x), z = 1, 2,..., N sp} and the corresponding model weight coefficients {α z , z = 1, 2,..., N sp}. The input weight matrix of the zth basic echo state network G n (x) of G n,z (x) is The output weight matrix is The sparse coding dimensionality reduction matrix is The corresponding hyperparameter group is {r z , L z , C z}; where, G n,z (x) represents the zth basic echo state network of the integrated echo state network prediction model G n (x) for the nth day; r z , L z , C z respectively represent the leakage rate, reservoir parameter, and regularization parameter of G n,z (x); let z = 1;

[0094] (5.2) Update the reservoir echo state variables of G n,z (x):

[0095]

[0096] Wherein, u(n) is the input vector on the nth day, and x z (n - 1) and x z (n) are the reservoir echo state variables of G n (x) of the zth basic echo state network G n,z (x) on the (n - 1)th day and the nth day respectively, and tanh is the hyperbolic tangent function;

[0097] (5.3) Reduce the dimension of the reservoir echo state variables of G n,z (x) to obtain the dimension-reduced reservoir echo state variable h z (n):

[0098]

[0099] (5.4) Output the prediction result y n,z (n) of G z (x):

[0100]

[0101] (5.5) If z < N sp , then let z = z + 1 and return to step (5.2); otherwise, calculate the load prediction result on the nth day

[0102]

[0103] Wherein, α p is the model weight coefficient of the pth basic echo state network of the integrated echo state network prediction model G n (x) on the nth day.

[0104] 6) After the nth day, record the actual load data on the nth day To avoid frequent retraining of the model to further improve the prediction speed, the Kalman filter algorithm with modified covariance is used to update the parameters of the integrated echo state network prediction model G n (x) on the nth day to obtain the integrated echo state network prediction model G n+1 (x) on the (n + 1)th day; including:

[0105] (6.1) Set the integrated echo state network prediction model G n (x) on the nth day consists of N spA basic echo state network {G n,z (x), z = 1, 2, …, N sp} and the corresponding model weight coefficients {α z , z = 1, 2, …, N sp} are integrated to obtain. The input weight matrix of the z-th basic echo state network G n (x) of G n,z (x) is The output weight matrix is The sparse coding dimensionality reduction matrix is The corresponding hyperparameter group is {r z , L z , C z}; where, G n,z (x) represents the z-th basic echo state network of the integrated echo state network prediction model G n (x) on the n-th day; r z , L z , C z respectively represent the leakage rate, reservoir parameter, and regularization parameter of G n,z (x); let z = 1;

[0106] (6.2) Calculate the weight component n,z of the m-th period of G (x) n,z For the covariance matrix of the prior estimation error of the weight component of the m-th period of G

[0107]

[0108]

[0109] In the formula, represents the m-th row, represents the transpose of the m-th row, P n-1,z (m), Q n-1,z (m) respectively represent the covariance matrix of the posterior estimation error and the process noise covariance matrix of the weight component of the m-th period of the z-th basic echo state network G n-1 (x) of the integrated echo state network prediction model G n-1,z (x) on the (n - 1)-th day;

[0110] (6.3) Calculate the Kalman gain K n,z (x) of the m-th period of G n,z (m):

[0111]

[0112] In the formula, h z (n) represents the reservoir echo state variable after dimensionality reduction of G n,z (x);

[0113] (6.4) Update the posterior estimation error covariance matrix P n,z (m) of G n,z (x) at the m-th time period:

[0114]

[0115]

[0116] In the formula, F e (m) is the correction factor of P n,z (m), and c(m) is the variance of the prediction error in the m-th time period from the (n - k)-th day to the n-th day, which is calculated as follows:

[0117]

[0118] In the formula, represents the actual load value at the m-th time period on the (n - j)-th day, represents the z-th basic echo state network G n-j (x) of the integrated echo state network prediction model G n-j,z (x) at the m-th time period;

[0119] (6.5) Update the weight component n,z of G (x) at the m-th time period to obtain the integrated echo state network prediction model G n+1 (x) of the (n + 1)-th day, where the z-th basic echo state network G n+1,z (x) has an output weight matrix

[0120]

[0121]

[0122] In the formula, represents the actual load value at the m-th time period on the n-th day, is the weight component n+1,z of G

[0123] (6.6) If z < N sp , let z = z + 1, and return to step (6.2); otherwise, obtain the integrated echo state network prediction model G n+1 (x) of the (n + 1)-th day.

[0124] 7) If c ≤ T L , then let n = n + 1, c = c + 1, and return to step 5); otherwise, proceed to step 8);

[0125] 8) Based on the prediction errors of each basic echo state network {G n (x), z = 1, 2, …, N n,z} of G sp (x) from the (n - T L + 1)-th day to the n-th day, using the Adaboost algorithm, recalculate the model weight coefficients {α n} of each basic echo state network {G n,z (x), z = 1, 2, …, N sp} of G z (x) to obtain new model weight coefficients {α′ z}, let n = n + 1, c = 1, and return to step 5); where G n,z (x) represents the z-th basic echo state network in the integrated echo state network prediction model g n (x) on the n-th day.

[0126] An adaptive integrated load forecasting method for a distribution network facing edge computing in the present invention uses a lightweight neural network ESN as a basic predictor and the Tradaboost algorithm as the main integration framework. By deleting ESN predictors with low prediction performance and correspondingly improving ESN predictors with high prediction performance, the integration scale is reduced, and the computational overhead and storage resources for subsequent predictions are reduced. In the rolling prediction stage of the day-ahead, combined with the actual load data of each day, the Kalman filtering algorithm with modified covariance is used to update each basic predictor; in the long-term prediction, the Adaboost algorithm is regularly used to redistribute the weights of each basic ESN predictor to solve the adaptive problem of the edge-side load prediction model.

[0127] The following gives a specific example:

[0128] To prove the effectiveness of the model (Model-4) of the present invention, a simulation experiment was carried out on the load dataset of the Global Energy Forecasting Competition (GEFCOM2012), and the results were mainly compared with the following models:

[0129] Conventional integrated ESN model (Model-1): Uses ESN as the basic predictor and Tradaboost as the integrated learning framework, and does not adjust the model parameters for subsequent predictions;

[0130] Sparse integrated ESN model (Model-2): Based on Model-1, automatically screen the basic predictors to reduce the final integration scale, and do not adjust the model parameters for subsequent predictions;

[0131] Sparse integrated ESN-KF model (Model-3): The initial model is the same as Model-2, and the Kalman filter algorithm with corrected covariance is used for model update every day.

[0132] In addition, it is also compared with traditional machine learning models and traditional ensemble learning models. The traditional machine learning model uses the Gradient boosting decision tree (GBDT) model, and the traditional ensemble learning model uses the widely used Adaboost ensemble algorithm.

[0133] The evaluation index uses the common Mean Absolute Percentage Error (MAPE), which is calculated as follows: The calculation method is as follows:

[0134]

[0135] In the formula, N is the total number of time periods for day-ahead prediction, x i is the actual load value at the i-th hour, the predicted load value at the i-th hour.

[0136] The method proposed in the present invention can be deployed on an actual edge computing device with only 1GB of memory and an 800MHz CPU. To facilitate comparison of the prediction performance and calculation speed of different models, all model construction and training are carried out on Matlab2018, and the hardware platform uses an Intel Core i7 CPU with a main frequency of 2.9GHz and 16GB of memory.

[0137] In Example 1, regions 1, 11, 12, 15, and 20 in the global energy forecasting competition load dataset are selected as the research objects. The training set uses the data between January 1, 2007 and March 31, 2007, and the test set uses the data between April 1, 2007 and April 7, 2007. Input relevant parameters: The number of samples for initializing the network is n0 = 20, the number of samples for training is Γ = 70, the candidate set of echo state network ESN leakage rates {0.92, 0.94, 0.96, 0.98}, the candidate set of reservoir parameters {700, 800, 900, 1000, 1100, 1200}, the candidate set of regularization parameters {10, 10 2 , 10 3 , 10 4}, the integration scale N0 = 80, the target low-dimensional dimension D L = 100, the number of days in the source domain Md = 50, the number of days in the target domain N d = 20, the number of correction factor sample calculations k = 30, the long-term prediction days threshold T L = 30, the main diagonal elements of the covariance matrix set {P} of the initial posterior estimation error are taken as 10 -2 , and the rest are 0; the main diagonal elements of the initial process noise covariance matrix set {Q} are taken as 10 -4 , and the rest are 0; the initial measurement noise covariance vector R M Each element is taken as 10 -4 .

[0138] All integrated prediction models are conducted 10 independent experiments, and the average value is taken as the final result. The prediction results of different models in each region are compared as shown in Table 1 in the appendix. The prediction results of different methods in Region 1 are as shown in the appendix Figure 2 shown. Since the number of days in the test set is only 7 days, the weights of the basic predictors are not adjusted. Therefore, the results of Model-4 are the same as those of Model-3, so the results of Model-4 are not listed. From Table 1 in the appendix and the appendix Figure 2 it can be seen that the prediction result of GBDT is the worst, and the result of Adaboost.R2 is the second. In Regions 1, 11, and 15, the prediction performance of Model-4 is the best, but in Regions 12 and 20, the best prediction performance is not achieved. This is due to the fact that the test set time is too short and the process noise matrix and measurement noise values are not ideal values

[0139] The comparison of the calculation times of different models is shown in Table 2 in the appendix. The average training time and average prediction time of GBDT are the shortest. This is because other models all adopt ensemble learning, and the time cost of training numerous basic predictors is large. Correspondingly, the prediction performance of GBDT is the worst; since Model-1 adopts the Tradaboost framework, the training time is slightly longer than that of Adaboost. Model-2 and Model-4 add an automatic screening mechanism, and the average training time is the longest, but the increased time cost is relatively low. In terms of the average prediction time, since Model-4 adopts Kalman filtering while other models do not need to update parameters, the average prediction time of Model-4 is the longest. However, in subsequent predictions, Model-4 does not need to be retrained, and the average prediction time is much less than the average training time. Overall, it still has advantages

[0140] To verify the adaptability of Model-3 and Model-4, experiments of Example 2 were conducted for Regions 12 and 20 in Example 1 where the best prediction results were not obtained in the short term. Other conditions in Example 2 were the same as those in Example 1, except that the test time was extended, and the test set time was from April 1, 2007 to January 31, 2008. The prediction performance gap between Model1-4 and GBDT was relatively obvious. For the sake of intuitive and clear comparison, only the daily moving average MAPE curves of Model1-4 in Region 20 from April 1, 2007 to May 31, 2007 are given, as shown in Figure 3 shown. The specific statistical errors are shown in Table 3 in the appendix. As can be seen from Figure 3 the appendix and Table 3 in the appendix, since Model-3 and Model-4 update the model parameters every day, even in the prediction for nearly a year, without retraining, the accuracy of the prediction results can be guaranteed. After one month, the prediction performance of Model-3 and Model-4 was stably better than that of Model-1 and Model-2. Since Model-4 redistributes the weights of the basic predictors every month, it still has a slight improvement compared with Model-3.

[0141] In summary, an adaptive integrated load forecasting method for distribution networks oriented to edge computing proposed by the present invention can effectively perform day-ahead forecasting on the edge-side load, with less computational overhead and storage resource occupancy, and has the feasibility of being deployed on edge devices.

[0142] Table 1 Comparison of prediction results of different methods in Example 1 (MAPE, %)

[0143] Region number GBDT Adaboost.R2 Model-1 Model-2 Model-4 1 25.09 24.86 13.31 13.30 12.85 11 25.15 23.54 10.38 10.38 10.31 12 25.09 24.34 8.04 8.05 9.32 15 22.12 22.02 14.88 14.88 14.76 20 19.43 20.08 11.02 11.02 12.05

[0144] Table 2 Comparison of calculation times of different methods in Example 1

[0145]

[0146] Table 3 Comparison of prediction results of different methods in Example 2 (MAPE, %)

[0147] Region number Model-1 Model-2 Model-3 Model-4 12 8.42 8.42 6.51 6.43 20 9.54 9.56 6.25 6.19

Claims

1. An adaptive integrated load forecasting method for a distribution network oriented to edge computing, characterized in that Including the following steps: 1) Input the historical load dataset E and the historical ambient temperature dataset T of the area under the jurisdiction of the edge computing device of the distribution network as the initial training set Ω; set the initial number of training samples n0, the number of training samples Γ, and the candidate set of leakage rates of the echo state network Candidate set of reservoir parameters Candidate set of regularization parameters Integration scale N0 = N1 × N2 × N3, low-dimensional dimension D of echo state variables L , number of days M in the source domain d , number of days N in the target domain d , set {P} of covariance matrices of initial posterior estimation errors, set {Q} of initial process noise covariance matrices, initial measurement noise covariance vector R M , number k of training samples for calculating correction factors, long-term prediction days threshold T L ; where N1, N2, and N3 are the total numbers of candidates in the candidate sets for the leakage rate, reservoir parameters, and regularization parameters, respectively; is the N1-th leakage rate in the leakage rate candidate set, is the N2-th reservoir parameter in the reservoir parameter candidate set, is the N3-th regularization parameter in the regularization parameter candidate set, and the integration scale N0 is the total number of basic echo state networks; 2) Perform min-max normalization on the load history data set E and the ambient temperature history data set T in the initial training set Ω respectively, and perform training sample division to obtain the training set Ω′; 3) Due to the limitations of the computing power of the edge computing device, the echo state network in the lightweight neural network is selected as the core prediction algorithm, and sparse coding is introduced to reduce the dimension of the reservoir echo state variables. Based on the training set Ω′, N0 basic echo state networks {G l (x), l = 1, 2, …, N0} are trained respectively; where G l (x) is the l-th basic echo state network; 4) Set the current prediction day number n = 1 and the long-term prediction day number identifier c = 1. To reduce the storage resource requirements and integration scale of the edge computing device, for the N0 basic echo state networks {G l (x), l = 1, 2, …, N0} obtained by training, use the Tradaboost algorithm with sparse adaptive boosting to automatically delete non-critical basic echo state networks, and for the remaining basic echo state networks {G 1,z (x), z = 1, 2, …, N sp} and the corresponding model weight coefficients {α z , z = 1, 2, …, N sp}, integrate to obtain the integrated echo state network prediction model G 1 (x) for the first day; Among them, G 1,z (x) is the z-th basic echo state network of the integrated echo state network prediction model G 1 (x) on the first day, and N sp is the number of remaining basic echo state networks after automatic screening; 5) Before the start of the current prediction day, that is, before the nth day, use the integrated echo state network prediction model G n (x) of the nth day to perform day-ahead prediction. Input the actual load value of the day before the prediction day, the actual load values of the 7 days before the prediction day, and the ambient temperature data of the weather forecast for the prediction day to obtain the load prediction result for the nth day 6) After the end of the nth day, record the actual load data of the nth day To avoid frequent retraining of the model to further improve the prediction speed, the Kalman filter algorithm with corrected covariance is used to update the parameters of the integrated echo state network prediction model G n (x) of the nth day to obtain the integrated echo state network prediction model G n+1 (x) of the (n + 1)th day; 7) If c ≤ T L , then let n = n + 1, c = c + 1, and return to step 5); otherwise, proceed to step 8); 8) Based on G n (x), for each of the basic echo state networks {G n,z (x), z = 1, 2, …, N sp} from the (n - T L +1)-th day to the n-th day, using the Adaboost algorithm, recalculate the model weight coefficients {α n} of each of the basic echo state networks {G n,z (x), z = 1, 2, …, N sp} to obtain the new model weight coefficients {α′ z}, let n = n + 1, c = 1, and return to step 5); where G z (x) represents the z-th basic echo state network in the integrated echo state network prediction model G n,z (x) on the n-th day. n (x).

2. The adaptive integrated load forecasting method for a distribution network oriented to edge computing according to claim 1, wherein, The training sample division described in step 2) to obtain the training set Ω′ is specifically: based on the normalized initial training set Ω, perform training sample division. The input data of the i-th training sample is: the actual load value 1 day before the i-th prediction day, the actual load values of the previous 7 days, and the ambient temperature data of the weather forecast on the i-th prediction day, and merge the above input data into a column vector u(i), and the dimension is denoted as M×1; The label of the i-th training sample is the actual load value y(i) on the i-th prediction day in the training set Ω′, and the label dimension of the training sample is denoted as N×1; where N is the total number of time periods for day-ahead prediction.

3. The adaptive integrated load forecasting method for a distribution network oriented to edge computing according to claim 1, wherein Step 3) Based on the training set Ω′, train N0 basic echo state networks {G l (x), l = 1, 2, …, N0} respectively, specifically as follows: (3.1) Permute the leakage rate candidate set the candidate set of reservoir parameters the candidate set of regularization parameters perform permutation and combination, and use them as the hyperparameters of N0 basic echo state networks; (3.2) Set the l-th basic echo state network G l (x) with the hyperparameter group {r l , L l , C l}, and initialize the reservoir echo state variable x l (0) = 0, and randomly generate the input weight matrix the internal connection matrix of the reservoir the sparse coding random weight matrix A l , the random bias matrix B l ; where r l , L l , C l represent the leakage rate, the reservoir parameter, and the regularization parameter of G l (x) respectively, x l has the dimension of L l ×1, has the dimension of L l ×M, has the dimension of L l ×L l , A l has the dimension of D L ×(M + L l ), B l has the dimension of D L ×Γ, M is the dimension of the input column vector u of a single training sample, and Γ is the number of training samples; (3.3) Based on the first n0 initial training samples in the training set Ω′, recursively calculate the reservoir echo state variables of the l-th basic echo state network G l (x), and the recurrence formula is as follows: where \(u(i)\) is the input vector of the \(i\)-th training sample in the training set \(\Omega'\), \(x\) l (i - 1)\(, x\) l (i)\( are respectively the reservoir echo state variables of the \((i - 1)\)-th and \(i\)-th training samples in the training set \(\Omega\) ′ , and \(\tanh\) is the hyperbolic tangent function; (3.4) Form: (M + L l ) × 1 - dimensional column vector s l (i) = [u(i); x l (i)], (M + L l ) × Γ - dimensional matrix S l = [s l (n0 + 1), s l (n0 + 2), …, s l (n0 + Γ)], N×Γ dimensional matrix Y l = [y(n0 + 1), y(n0 + 2), …, y(n0 + Γ)], where y(n0 + 1) is the actual load value of the (n0 + 1)-th training sample in the training set Ω ′ ; Calculate the low-dimensional feature matrix H l , sparse coding dimensionality reduction matrix Output weight matrix Complete the training of the l-th basic echo state network G l (x): H l = A l S l + B l (2) where I is the identity matrix, H l is of dimension D L × Γ, is of dimension D L × (M + L l ), is of dimension N × D L ; (3.5) Repeat steps (3.2) to (3.4) until the training of N0 basic echo state networks is completed.

4. A method for adaptively integrating load forecasting of a distribution network for edge computing according to claim 1, characterized in that, The Tradaboost algorithm using sparse adaptive boosting in step 4) automatically deletes non-critical basic echo state networks, and the remaining basic echo state networks {G 1,z (x), z = 1, 2, …, N sp} and the corresponding model weight coefficients {α z , z = 1, 2, …, N sp} are integrated to obtain the integrated echo state network prediction model G 1 (x) on the first day, specifically including: (4.1) Divide the Γ training samples for training into a source domain data set consisting of the first M training samples and a target domain data set consisting of the remaining N training samples in chronological order; Let the source domain weight parameter d and the current iteration number l = 1, and calculate the initial weights of each training sample in the training set Ω′ d (4.2) For the training set Ω′ with the sample weight distribution D l-1 calculate the maximum regression error E l of the l-th basic echo state network G l : E l = max(∑|y(i) - G l (u(i))|), i = 1, 2, …, M d + N d (6) where G l (u(i)) represents the predicted value of the l-th basic echo state network G l (x) for the i-th training sample in the training set Ω′, y(i) is the actual value for the i-th training sample in the training set Ω′, G l (u(i)) and y(i) are both of dimension N×1, and ∑|y(i) - Gl(ui)| represents the sum of the absolute values of all elements of the vector yi - Gl(ui); (4.3) Calculate the l-th basic echo state network G l (x) The relative regression error e of each training sample on the training set Ω′ l,i : (4.4) Calculate the l-th basic echo state network G l (x) The overall regression error e on the training set Ω′ l and the target domain weight parameter β l : In the formula, represents taking the relatively smaller value between and 0.5, and represents the weight of the i-th training sample in the l-th iteration training set Ω'. (4.5) Calculate the model weight coefficient α l of the l-th basic echo state network G l and the cumulative model weight coefficient R: where α v represents the model weight coefficient of the v-th selected basic echo state network G v (x) at the v-th iteration; (4.6) If l = 1, then set l = l + 1 and return to step (4.2); otherwise, find two basic echo state networks at the h1-th and h2-th iterations that need to adjust the model weight coefficients according to the following formula, denoted as and Then proceed to step (4.7): where \(u(i)\) represents the input vector of the \(i\)-th training sample in the training set \(\Omega'\), \(G\) h \((u(i))\) represents the predicted value of the \(h\)-th selected basic echo state network \(G\) h \((x)\) for the \(i\)-th training sample \(u(i)\) in the training set \(\Omega'\), \(G\) v \((u(i))\) represents the predicted value of the \(v\)-th selected basic echo state network \(G\) v \((x)\) for the \(i\)-th training sample \(u(i)\) in the training set \(\Omega'\), represents the Hadamard product, that is, the corresponding elements of the matrices are multiplied; (4.7) Calculate the transfer step size p of the optimal model weight coefficient * And make the corresponding adjustment to the weight coefficient of the basic echo state network model: In the formula, represents the predicted value of the i-th training sample in the training set Ω′, represents the predicted value of the i-th training sample in the training set Ω′, represents the model weight coefficient of represents the model weight coefficient of; (4.8) If then the cumulative model weight coefficient R is adjusted using the following formula (16), the weight coefficient of the basic echo state network model with a model weight coefficient less than 0 is set to 0, and the basic echo state network is deleted: (4.9) Update the training sample weights of the source domain and the target domain: (4.10) If l < N0, then let l = l + 1 and return to step (4.2); otherwise, calculate the integrated echo state network prediction model G 1 (x) for the first day:

5. The adaptive integrated load forecasting method for a distribution network oriented to edge computing according to claim 1, characterized in that Step 5) includes: (5.1) Set the integrated echo state network prediction model \(G\) for the \(n\)th day n (x) is composed of sp \(N\) n,z basic echo state networks \(\{G\) sp (x), z = 1, 2, …, \(N\) z \}\) and the corresponding model weight coefficients \(\{\alpha\) sp , z = 1, 2, …, \(N\) n \}\). The input weight matrix of the \(z\)th basic echo state network \(G\) n,z (x) of \(G(x)\) is The output weight matrix is The sparse coding dimensionality reduction matrix is The corresponding hyperparameter group is \(\{r\) z , \(L\) z , \(C\) z \}\); where \(G\) n,z (x) represents the \(z\)th basic echo state network of the integrated echo state network prediction model \(G\) n (x) for the \(n\)th day; \(r\) z , \(L\) z , \(C\) z respectively represent the leakage rate, reservoir parameter and regularization parameter of \(G\) n,z (x); let \(z = 1\); (5.2) Update the reservoir echo state variable of G n,z (x): Where u(n) is the input vector of the nth day, x z (n-1), x z (n) are G n The zth basic echo state network G of (x) n,z (x) the echo state variable of the reservoir on day n-1 and day n, tanh is the hyperbolic tangent function; (5.3) for G n,z The dimension of the reserve pool echo state variable (x) is reduced to obtain the reduced-dimensional reserve pool echo state variable h z (n): (5.4) Output G n,z (x) Predicted result y z (n): If z < N sp , then let z = z + 1, and return to step (5.2); otherwise, calculate the load forecasting result for the nth day where α p is the model weight coefficient of the p-th basic echo state network of the integrated echo state network prediction model G n (x) on the n-th day.

6. The adaptive integrated load forecasting method for a distribution network oriented to edge computing according to claim 1, wherein, Step 6) includes: (6.1) Set the integrated echo state network prediction model G for the nth day n (x) is composed of sp N n,z basic echo state networks {G sp (x), z = 1, 2, …, N z} and the corresponding model weight coefficients {α sp , z = 1, 2, …, N n . The input weight matrix of the zth basic echo state network G n,z (x) of G is The output weight matrix is The sparse coding dimensionality reduction matrix is z The corresponding hyperparameter group is {r z , L z , C n,z}; where G n (x) represents the zth basic echo state network of the integrated echo state network prediction model G z (x) for the nth day; r z , L z , C n,z represent the leakage rate, reservoir parameter, and regularization parameter of G (x) respectively; let z = 1; (6.2) Calculate G n,z The weight component of the m-th period of G For G n,z The covariance matrix of the prior estimation error of the weight component of the m-th period of G Perform an update: In the formula, denotes the m-th row, denotes the transpose of the m-th row, P n-1,z (m), Q n-1,z (m) respectively denote the covariance matrix of the posterior estimation error of the z-th basic echo state network G n-1 (x) and the process noise covariance matrix of the m-th time period of the echo state network G n-1,z (x); (6.3) Calculate G n,z The Kalman gain K at the m-th time period of n,z (m): where h z (n) represents G n,z (x) the reservoir echo state variable after dimensionality reduction; (6.4) Update G n,z The posterior estimation error covariance matrix P of the m-th period of (x) n,z (m): where F e (m) is the correction factor of P n,z (m), and c(m) is the variance of the prediction error in the m-th period from the (n - k)-th day to the n-th day, calculated as follows: In the formula, represents the actual load value at the m-th time period on the (n - j)-th day, represents the load prediction value at the m-th time period of the z-th basic echo state network G n-j (x) of the integrated echo state network prediction model G n-j,z on the (n - j)-th day; (6.5) Update G n,z The weight component of (x) in the m-th time period Obtain the integrated echo state network prediction model G for the (n + 1)-th day n+1 The z-th basic echo state network G of (x) n+1,z The output weight matrix of (x) Wherein, represents the actual load value at the m-th time period on the n-th day, is the weight component of the m-th time period of G n+1,z (x); If z < N sp , let z = z + 1, and return to step (6.2); otherwise, obtain the integrated echo state network prediction model G n+1 (x) for the (n + 1)-th day.

Images