CN106022521A - Hadoop framework-based short-term load prediction method for distributed BP neural network - Google Patents

Abstract

The invention discloses a short-term load forecasting method of a distributed BP neural network based on Hadoop architecture, specifically: obtaining an initial load data set; splitting the load data set into small data sets, and storing them in data nodes of a distributed file system Middle; initialize the BP neural network parameters, and upload the parameter set to the distributed file system; train the BP neural network according to the current load sample, and obtain the correction value of the weight and threshold of the BP neural network under the current data set; according to the key The key value of the value pair counts the sum of the weights and threshold parameters of each layer of the network and between layers; judge whether the current iteration task has reached the convergence accuracy or the maximum number of iterations, and if so, establish a distributed BP neural network model , if not, modify the weights and threshold parameters of the BP neural network; input the forecast day data to obtain the load power data of the forecast day. The invention improves the speed of load forecasting and meets the requirement of load forecasting accuracy.

Description

Short-term load forecasting method based on distributed BP neural network based on Hadoop architecture

technical field

The invention relates to the technical field of short-term load forecasting application in the scene of combining power system and big data, and in particular to a short-term load forecasting method based on a distributed BP neural network of Hadoop architecture.

Background technique

Power load forecasting is of great significance in ensuring power system planning, reliable and economical operation. With the continuous progress of modern technology and the deepening of smart grid, the theory and technology of load forecasting have been greatly developed. Over the years, power load forecasting methods and theories have emerged continuously. Technologies such as time series method, fuzzy theory, regression analysis method, regression support vector machine, Bayesian and neural network have provided good technical support for power load forecasting.

Existing algorithms still have certain limitations. Time series method: It has high accuracy for historical data, and is not sensitive to weather factors in short-term load forecasting, so it is difficult to solve the problem of low accuracy of short-term load forecasting caused by meteorological factors. Regression analysis method: Quantitatively describe the quantitative relationship between observed variables from the perspective of statistical average significance, but it is very limited by the scale of load data. Regression support vector machine: This method has good generalization ability, but at the same time, the training time is too long due to the optimization of the penalty coefficient c, the e value of the loss function and the gamma value of the kernel function, especially in The larger the training sample set, the more prominent it is.

With the emergence of massive data on smart power consumption, the existing forecasting algorithms cannot meet the requirements of forecasting speed and forecasting accuracy. Therefore, it is necessary to find a new method that can meet the analysis of massive power consumption big data. BP neural network has strong nonlinear mapping ability, self-learning ability and fault-tolerant ability. When it is applied to small load data sets, it has the advantages of high prediction accuracy and fast training speed. However, since the algorithm will conduct a round of training for each load input and output sequence to calculate and obtain the correction amount of the weights and thresholds of each layer of the network, when the amount of data is very large, the amount of calculation will become very large, and the single-machine serial training time It will be possible to reach several hours, or even greater. Therefore, the problem of short-term load forecasting based on massive data is still an urgent problem to be solved.

Contents of the invention

The technical problem to be solved by the present invention is to provide a short-term load forecasting method based on the distributed BP neural network of the Hadoop architecture. The method is based on the Hadoop cluster platform, and exerts the powerful nonlinear mapping ability of the BP neural network to improve the load forecasting speed. Meet the requirements of load forecasting accuracy.

In order to solve the problems of the technologies described above, the technical solution adopted in the present invention is:

A kind of short-term load forecasting method of distributed BP neural network based on Hadoop framework, comprises the following steps:

Step 1: Obtain the initial load data set;

Step 2: According to the MapReduce framework of the Hadoop architecture, split the load data set into small data sets and store them in the data nodes of the distributed file system; the small data sets are represented by key-value pairs <key, value>, and the key value is the offset of the first character of the line relative to the first address of the text file, and the value value will be parsed into the weights and thresholds of each layer of the current BP neural network;

Step 3: Initialize the parameters of the BP neural network, including the number of layers of the input layer, hidden layer and output layer, the weight between the input layer and the hidden layer, the threshold value of the hidden layer neurons, the distance between the hidden layer and the output layer The weights between the neurons in the output layer and the threshold of the neurons in the output layer are uploaded to the distributed file system;

Step 4: In the Map stage, read the parameters in the distributed file system, including weights and thresholds, and restore the BP neural network at the beginning of each subtask; perform normalization of the input signal of the BP neural network according to the data assigned by the subtasks. To the backpropagation of the transfer and error signals, obtain the weights and threshold corrections of the BP neural network under the current data set, and use the key-value pairs as the input parameters of the Reduce stage;

Step 5: In the Reduce phase, after the BP neural network trains all data sets, according to the key value in the key-value pair <key, value> that identifies the corresponding weights and thresholds of neurons in the input layer, hidden layer, and output layer, statistics The influence of all load data samples on the weights and thresholds of each neuron after training is completed, and the results are output to the distributed file system;

Step 6: Determine whether the convergence accuracy or the maximum number of iterations is reached under the current iteration task; if so, according to the number of layers of the input layer, hidden layer and output layer of the BP neural network, and the weights and thresholds in the distributed file system Parameters to establish a distributed BP neural network model, if not, modify the weights and threshold parameters of the BP neural network;

Step 7: According to the distributed BP neural network model, input the forecast day data for forecasting, and obtain the load power data of the forecast day.

According to the above-mentioned scheme, in the step 4, calculate and obtain the weight of the BP neural network, the correction amount of the threshold under the current data set, specifically:

Δw′ _ki (τ+1)=(1-ρ)ηΔw _ki (τ+1)+ρΔw _ki (τ),

Δ′α _k (τ+1)=(1-ρ)ηΔα _k (τ+1)+ρΔα _k (τ),

Δ′w _ij (τ+1)=(1-ρ)ηΔw _ij (τ+1)+ρΔw _ij (τ),

Δ′θ _i (τ+1)=(1-ρ)ηΔθ _i (τ+1)+ρΔθ _i (τ),

Among them, Δw' _ki is the final weight correction value between the kth node in the output layer and the i-th node in the hidden layer; Δ'α _k is the final threshold correction value of the kth node in the output layer; Δ'w _ij represents the final weight correction amount between the i-th node in the hidden layer and the j-th node in the input layer; Δ′θ _i is the final threshold value correction amount of the i-th node in the hidden layer; ρ is the momentum factor; τ is number of iterations.

According to the above scheme, it also includes improving the distributed BP neural network model, that is, introducing a momentum factor, and improving the distributed BP neural network model by means of multiple calculations and averaging.

Compared with the prior art, the invention has the beneficial effects of adopting the Hadoop cluster platform for effectively processing massive big data, exerting the powerful nonlinear mapping capability of the BP neural network, increasing the speed of load forecasting, and meeting the requirements of load forecasting accuracy.

Description of drawings

Figure 1 is a schematic diagram of a typical three-layer BP neural network structure.

Fig. 2 is a schematic structural diagram of a distributed BP neural network prediction model in the present invention.

Fig. 3 is the topological diagram of the laboratory Hadoop cluster platform in the present invention.

Fig. 4 is a comparison chart of the MapReduce-BP load prediction result in the present invention and the traditional BP prediction result.

detailed description

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. The present invention has three characteristics: one, fully analyze the forward transmission of the input signal of the traditional BP neural network, and the reverse propagation process of the error signal; two, combine the training process of the BP neural network with the MapReduce framework, research and realize it by Java language The distributed BP neural network model based on the MapReduce framework, hereinafter referred to as the MapReduce-BP model; 3. The momentum factor is introduced, and the method of calculating the average value for multiple times is used to improve the problem that the BP neural network is easy to fall into local convergence and improve its anti-oscillation ability. The details are as follows:

1. Analysis of traditional BP neural network prediction principle

1) Basic model of BP neural network

In 1986, scientists headed by Rumelhart and McCelland proposed the BP neural network, which is a multi-layer feed-forward neural network that can learn and store a large number of input-output pattern mapping relationships without revealing the mathematical equations of the mapping relationship in advance. , which consists of an input layer, a hidden layer and an output layer. Figure 1 is a typical structural diagram of a three-layer BP neural network. Layers are fully interconnected. There is no interconnection between the same layer. The hidden layer can be one or more layers. In Figure 1, x _j represents the input of the j-th node in the input layer; w _ij represents the weight between the i-th node in the hidden layer and the j-th node in the input layer; θ _i is the weight of the i-th node in the hidden layer Threshold; φ is the activation function of the hidden layer; w _ki represents the weight between the kth node in the output layer and the i-th node in the hidden layer; α _k is the threshold of the kth node in the output layer; ψ is the output layer The activation function of ; ok represents the output of the _kth node.

2) BP neural network signal transmission and error correction

The basic BP neural network algorithm consists of two parts: the forward transmission of the signal and the backpropagation of the error. That is, the calculation of the actual output is carried out in the direction from input to output, and the correction process of the weights and thresholds of each layer is from the direction of output to input. conduct. According to the parameters shown in Figure 1, the output signal, weights and thresholds of each layer of the BP neural network are calculated and adjusted.

(1) The forward transfer process of the input signal

According to the structure diagram of BP neural network in Figure 1, we can know the input net _i and output o _i of the i-th node of the concerned hidden layer, the input net _k and output o _k of the k-th node in the output layer respectively

${\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; o</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; q</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; o</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&psi;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&psi;</span>}<span\; class="notranslate">\; \&lsqb;</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; q</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; \&phi;</span>}<span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; \&rsqb;</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

(2) The backpropagation process of the error signal

Firstly, the output error of neurons in each layer is calculated layer by layer from the output layer, and then the weights and thresholds of each layer are adjusted according to the error gradient descent method, so that the final output of the adjusted network map can be close to the expected value. According to the error gradient descent method, correct the hidden layer to output layer weight correction Δw _ki , the output layer threshold correction Δα _k , the input layer to the hidden layer weight correction Δw _ij and the hidden layer threshold correction Δθ _i , as shown in Equation (5)-Equation (8), where η is the learning rate; P is the total number of training samples.

${\mathrm{<span\; class="notranslate">\; \&Delta;w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; P</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; o</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; o</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; \&Delta;\&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; P</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; o</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; \&Delta;w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; P</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; o</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; \&Delta;\&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; P</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; o</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&phi;</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; net</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

2. Realization of distributed BP neural network based on MapReduce framework

The update of BP neural network weights and thresholds has no direct relationship with the input order of samples. Therefore, the training process of BP neural network can adopt the method of dividing the training data into different PCs in the cluster, and each PC trains for different data. This method is very similar to the principle of the MapReduce programming framework. To this end, combined with the parallel characteristics of MapReduce framework and BP neural network data, a distributed prediction model of BP neural network based on MapReduce framework is established, which mainly includes three parts: Map stage, Reduce stage and driving function.

1) Map stage

Divide the text data set uploaded to the distributed file system HDFS and store the weights and thresholds of each layer of the BP neural network into several data subsets, and the data subsets are represented by key-value pairs <key, value>. The key value is the offset of the first character of the line relative to the first address of the text file, and the value value will be parsed into the weights and thresholds of each layer of the current BP neural network. In this stage, the forward transmission of the input signal and the back propagation of the error signal are carried out for the current load sample to obtain the correction value of the weight and threshold. The weight and threshold correction here correspond to the value of the output key-value pair in the Map stage, and the key value is the alias of the weight and threshold of each layer. The key points of the Map stage for generating weights and threshold corrections are as follows:

Input: the current load data sample and the weights and thresholds of each layer of the BP neural network.

Output: the weight of the current sample and the correction amount of the threshold.

Method: According to the forward propagation of the input signal and the reverse propagation of the error signal, calculate the correction value of the weight and threshold, as follows:

(1) setup function

According to the text of storing weights and thresholds in the HDFS file system, the weights (w _ki and w _ij ) and thresholds (α _k and θ _i ) of each layer of the current BP neural network are analyzed according to the storage principle and order.

(2) Correction of weight and threshold

After the momentum factor is introduced, the update formulas of the weights and thresholds of each layer will be transformed into Equation (9)-Equation (12). Among them, Δw' _ki is the final weight correction value between the kth node in the output layer and the i-th node in the hidden layer; Δ'α _k is the final threshold correction value of the kth node in the output layer; Δ'w _ij is the final weight correction amount between the i-th node in the hidden layer and the j-th node in the input layer; Δ′θ _i is the final threshold value correction amount of the i-th node in the hidden layer; ρ is the momentum factor; τ is number of iterations. The updated adjustment formula is as follows:

Δw′ _ki (τ+1)=(1-ρ)ηΔw _ki (τ+1)+ρΔw _ki (τ) (9)

Δ′α _k (τ+1)=(1-ρ)ηΔα _k (τ+1)+ρΔα _k (τ) (10)

Δ′w _ij (τ+1)=(1-ρ)ηΔw _ij (τ+1)+ρΔw _ij (τ) (11)

Δ′θ _i (τ+1)=(1-ρ)ηΔθ _i (τ+1)+ρΔθ _i (τ) (12)

The forward transmission of the load input signal and the backpropagation of the error signal can be completed through formula (1)-(4), formula (9)-(12), respectively, so as to obtain the current load sequence, each of the BP neural network The weight of the layer and the correction amount of the threshold.

(3) map function

The map function mainly combines parts (1) and (2). After completing the calculation of weights and threshold corrections, the context mode is used to output the weights and threshold corrections as key-value pairs <key,value>, the specific form is < IntWritable, Text>.

2) Reduce stage

According to the key value in the key-value pair, combined with the value, the statistics of the overall weight of the BP neural network and the update amount of the threshold are completed. Pay attention to two points: 1. At this time, the value form is an Iterable<Text> collection, and the collection dimension is the total number of load sequence training samples. Therefore, it is necessary to traverse step by step to obtain the correction amount string Text of the current sample before correction; 2. This stage is completed The sum of neural network weights and threshold corrections is the accumulation of numerical values. Therefore, the string text representing the corrections of weights and thresholds needs to be parsed into numerical form. The key points of the Reduce phase of summing statistical weights and threshold corrections are as follows:

Input: The key-value pair <IntWritable, Text> representing the single load sequence sample pair BP neural network weights and threshold-adjusted correction value output by the Map stage.

Output: The update amount of the overall weight and threshold of the BP neural network.

Method: According to the aliases of the weights and thresholds of each layer of the BP neural network, which is also the key of the input key-value pair at this stage, the update amounts of the overall weights and thresholds are accumulated and counted respectively. The important point of this stage is that in the reduce function, the traversal of the value of the input key-value pair, that is, Iterable<Text> is completed successively, and the weight value of the current sample, the threshold value correction string and the conversion of the Text text into a numerical form are obtained. , complete the accumulation of corresponding weights and threshold corrections according to the key value.

3) Driver function

The driver function can be understood as the configuration file of the whole program, which mainly completes the relevant settings of the job. In the driving function of the present invention, three related settings are mainly completed: one, generate the weight and threshold text of the initial BP neural network, and upload it to the HDFS distributed file system; two, create a MapReduce job, and set the Mapper class The map function, the data type of the output key-value pair <key, value>, and the reduce function of the Reducer class, the data type of the output key-value pair <key, value>; 3. Considering that the BP neural network is an iterative solution process , so the corresponding job type is an iterative MapReduce computing task, so it is necessary to set the iteration end standard, that is, set the maximum number of iterations and the error tolerance for training.

The technical solutions and technical effects of the present invention will be further described below through specific examples.

1. Calculation system and data processing

The data of the present invention come from load data and weather data collected by an actual power grid. The sampling time interval of each device is 1h, and the weather information includes dry bulb temperature and dew point temperature. Although the amount of laboratory data has not reached the scale of big data, the experimental data can be used to verify the correctness of the method of the present invention, thereby providing a new method for load forecasting in a big data environment. The training range is the electricity consumption data from January 1, 2014 to March 31, 2014. The forecast date is the power load at different times on April 1, 2014, as shown in Table 1. The research on the load data found that these data present the characteristics of continuity, periodicity and correlation. According to these characteristics and the research results of a large number of literatures, the sample attributes are determined as the load at the same time two weeks before the day, the load at the same time one week before the day, and the load at the same time two days before the day. Load at the same moment of the day, load at the same moment the day before, dry bulb temperature at the same moment the day before, dew point temperature at the same moment the day before, predicted dry bulb temperature at the same moment of the day and dew point temperature at the same moment of the day, predicted actual load at the same moment of the day, and its samples The data are shown in Table 2.

Table 1 Actual load data on April 1, 2014

Moment/h Load/MW Moment/h Load/MW Moment/h Load/MW 1 11483 9 15870 17 14508 2 10924 10 15965 18 14332 3 10711 11 15978 19 14219 4 10728 12 15823 20 14702 5 11027 13 15556 twenty one 15265 6 12128 14 15388 twenty two 14557 7 14043 15 15060 twenty three 13416 8 15413 16 14761 twenty four 12135

Table 2 Load training data sample set

Attributes value Load at the same time for the previous two weeks 11600MW Load at the same time in the previous week 11857MW Load at the same time for the two days before 11462MW Load at the same time the previous day 11203MW Dry bulb temperature at the same time the previous day 46°C The dew point temperature of the day before at the same time 43°C Forecast dry bulb temperature at the same moment of the day 41°C Predict the dew point temperature at the same time of the day 18°C Forecast the actual load at the same moment of the day 11483MW

It is worth noting that the BP neural network is sensitive to numbers between 0 and 1. Therefore, before inputting the original load sequence into the distributed BP neural network model, the data needs to be normalized. Afterwards, the denormalization process is performed to obtain the actual load forecast value.

2. Experimental results and analysis

This experiment is to compare the MapReduce-BP neural network with the traditional BP neural network algorithm to confirm two points. First, the correction process of the BP neural network weight and threshold is not related to the training sequence of the load input sequence, that is, BP The neural network training process can be transformed into data parallelism; secondly, the correctness of the idea of writing BP neural network load forecasting model based on the MapReduce framework is verified. Fig. 4 is a comparison chart of the MapReduce-BP load prediction result in the present invention and the traditional BP prediction result. It can be seen that the forecast curve based on the parallel MapReduce-BP neural network for the forecasted daily load is relatively consistent with the forecasted daily actual load curve, and is similar to the forecast effect of the traditional BP neural network forecast value. Among them, the average relative error and root mean square error of MapReduce-BP load forecasting are 3.95% and 1.97% respectively; the average relative error and root mean square error of traditional BP neural network forecasting are 3.92% and 1.93%. Thus, the correctness of the proposed MapReduce-BP load distribution forecasting model is confirmed.

Claims (3)

1. a short-term load forecasting method based on the distributed BP neural network of Hadoop framework, is characterized in that, comprises the following steps: Step 1: Obtain the initial load data set; Step 2: According to the MapReduce framework of the Hadoop architecture, split the load data set into small data sets and store them in the data nodes of the distributed file system; the small data sets are represented by key-value pairs <key, value>, and the key value is the offset of the first character of the line relative to the first address of the text file, and the value value will be parsed into the weights and thresholds of each layer of the current BP neural network; Step 3: Initialize the parameters of the BP neural network, including the number of layers of the input layer, hidden layer and output layer, the weight between the input layer and the hidden layer, the threshold value of the hidden layer neurons, the distance between the hidden layer and the output layer The weights between the neurons in the output layer and the threshold of the neurons in the output layer are uploaded to the distributed file system; Step 4: In the Map stage, read the parameters in the distributed file system, including weights and thresholds, and restore the BP neural network at the beginning of each subtask; perform normalization of the input signal of the BP neural network according to the data allocated by the subtasks. To the backpropagation of the transfer and error signals, obtain the correction amount of the weight and threshold of the BP neural network under the current data set, and use the key-value pair form as the input parameter of the Reduce stage; Step 5: In the Reduce phase, after the BP neural network trains all data sets, according to the key value in the key-value pair <key, value> that identifies the corresponding weights and thresholds of neurons in the input layer, hidden layer, and output layer, statistics The influence of all load data samples on the weights and thresholds of each neuron after training is completed, and the results are output to the distributed file system; Step 6: Determine whether the convergence accuracy or the maximum number of iterations is reached under the current iteration task; if so, according to the number of layers of the input layer, hidden layer and output layer of the BP neural network, and the weights and thresholds in the distributed file system Parameters, establish a distributed BP neural network model, if not, modify the weights and threshold parameters of the BP neural network; Step 7: According to the distributed BP neural network model, input the forecast day data for forecasting, and obtain the load power data of the forecast day. 2. the short-term load forecasting method of the distributed BP neural network based on Hadoop architecture as claimed in claim 1, is characterized in that, in described step 4, calculates and obtains the weight of BP neural network, the threshold value under the current data set Corrections, specifically: Δw′ _ki (τ+1)=(1-ρ)ηΔw _ki (τ+1)+ρΔw _ki (τ), Δ′α _k (τ+1)=(1-ρ)ηΔα _k (τ+1)+ρΔα _k (τ), Δ′w _ij (τ+1)=(1-ρ)ηΔw _ij (τ+1)+ρΔw _ij (τ), Δ′θ _i (τ+1)=(1-ρ)ηΔθ _i (τ+1)+ρΔθ _i (τ), Among them, Δw′ _ki is the final weight correction amount between the kth node in the output layer and the i-th node in the hidden layer; Δ′α _k is the final threshold correction amount of the kth node in the output layer; Δ′w _ij is the final weight correction amount between the i-th node in the hidden layer and the j-th node in the input layer; Δ′θ _i is the final threshold value correction amount of the i-th node in the hidden layer; ρ is the momentum factor; τ is the iteration frequency. 3. the short-term load forecasting method of the distributed BP neural network based on Hadoop architecture as claimed in claim 1 or 2, is characterized in that, also comprises improving distributed BP neural network model, promptly introduces momentum factor, adopts multiple times The method of computing the mean value is improved on the distributed BP neural network model.