Research on Power Load Forecasting Model Based on Hybrid Algorithm Optimizing BP Neural Network

: Short time load forecasting is essential for daily planning and operation of electric power system. It is the important basis for economic dispatching, scheduling and safe operation. Neural network, which has strong nonlinear fitting capability, is widely used in the load forecasting and obtains good prediction effect in nonlinear chaotic time series forecasting. However, the neural network is easy to fall in local optimum, unable to find the global optimal solution. This paper will integrate the traditional optimization algorithm and propose the hybrid intelligent optimization algorithm based on particle swarm optimization algorithm and ant colony optimization algorithm (ACO-PSO) to improve the generalization of the neural network. In the empirical analysis, we select electricity consumption in a certain area for validation. Compared with the traditional BP neutral network and statistical methods, the experimental results demonstrate that the performance of the improved model with more precise results and stronger generalization ability is much better than the traditional methods.


INTRODUCTION
Power load forecasting is an important part of the rapid development of electric power and relate to the long-term planning of power industry.At the same time, it is also an important part of power system planning and scheduling and is an important foundation for the economic operation of electric power system.According to the time classification, forecasting load can be classified as long-term, mediumterm, short-term and ultra-short term load forecasting [1].
At present, there are many kinds of methods to study power load forecasting.Each method has its advantages and disadvantages and there hasn't occurred a universal method for various situations yet.Considering different economic background and influence factors, we can choose different forecasting methods [2].According to advantages and shortcomings of each kind of electric power load forecasting method, the selection of reasonable forecasting model can improve the accuracy of power load forecasting and provide better decision-making basis for power grid.
Forecasting for the future development of less than 1 year is called short-term prediction.Many scholars and experts have carried out a theoretical study and practical simulation of a large number of short term load forecasting.The methods used in the prediction include regression models, time series model, and later to the intelligent models, such as neural network [3][4][5], support vector machine [6] and so on.Neural network has strong nonlinear fitting capability, it can be mapped to arbitrary complex nonlinear relationship through training samples, and can also intelligently adapts to arbitrary nonlinear variation in the short term.Its learning rule is simple, and easy to operate and implement.Therefore, this model can effectively meet the demand forecasting including arbitrary power economic indicators affected by multiple factors in the short term.However, the neural network is easy to fall into local optimum and appears the disadvantages of overfitting, low efficiency of calculation and poor generalization, which is difficult to guarantee the prediction accuracy of models [7][8].The methods commonly used for the parameter optimization of the neural network include: particle swarm algorithm, ant colony algorithm, genetic algorithm and so on.As the most commonly used swarm intelligence optimization algorithms, ant colony optimization and particle swarm optimization have greater optimization features.Ant colony optimization, which is the simulation of ant colony foraging process, has been successfully applied to many discrete optimization problems.Particle swarm optimization, which is the simulation of birds foraging process, is an efficient parallel search algorithm in continuous optimization field.Ant colony optimization uses pheromone to transmit information, while particle swarm optimization uses three information of the information of its own, individual extreme information, and global extreme information to guide the particle to the next iteration [9].Using the organic combination of the positive feedback principle and some heuristic algorithms, ant colony optimization is easy to run into prematurity and fall into local optimum.The organic combination of those two algorithms can overcome the shortcomings of them effectively and improve the computational efficiency significantly [10].
Given this, according to complementary characteristics of each method, this paper presents a hybrid optimization algorithm to optimize the parameters of neural network.The paper is organized as follows: The second chapter introduces the basic theory of neural network and optimization algorithm; the third chapter carries on the short-term load forecasting of a certain area using the proposed new method., at the same time, using the traditional BP algorithm and multiple linear regression method to predict the same target for the verification of accuracy and generalization ability the proposed model; the fourth chapter carries on the summary, and puts forward the further research work.

Neural network prediction model and BP learning algorithm
Neural network has strong nonlinear fitting capability, it can be mapped to arbitrary complex nonlinear relationship through training samples, and can also intelligently adapts to arbitrary nonlinear variation in the short term.Its learning rule is simple, and easy to operate and implement.Therefore, this model can effectively meet the demand forecasting including arbitrary power economic indicators affected by multiple factors in the short term.
A simple neural network is usually divided into three layers, including the input layer, hidden layer and output layer.Suppose the number of input layer neurons is N, number of hidden layer neurons is L, twice the number of input layer neurons according to the experience, that is L > 2N; the number of output layer neurons is M. The topological structure of BP neural network is shown in Fig. (1).
We put the BP neural network training samples in the input layer and output layer, then get the mapping from input to output.The mapping relationships don't need a specific calculation formula, which is adjusted by back propagation error signal neuron threshold and link weights, making the error minimum [11].
The error back propagation algorithm of the neural network is to define an error E r (typically a certain norm be- tween the output and the expected results), and then calculate the weight vector meeting the minimal error.If the error is seen as a continuous function (functional), it will meet that the partial derivative of each component of the weight vector is zero, but in fact, it is discrete, we need to use iteration to find the minimum gradient.
In practice, given a maximum number of iterations n and an error limit E l , using the error back propagation algorithm, every step of weight correction can make E r reducing, namely component of the weight vector go along the direction of the gradient decreases in advance.Although, when sample is large enough and n tends to infinite, the error will be convergence theoretically.Actually when the number of iterations reach n, the error E r may be still greater than the error limit E l .That means the weight vector does not meet the requirements, so network training fails.Of course, we also can use gradient limit as the terminating condition.In this case, non-convergence means gradient is no less than a certain value in iteration after n times, and thus the weight vector does not meet the requirements.Therefore, convergence is the weight vector of the meet gradient limit.Therefore, after setting the convergence conditions in the BP neural network, such as the relative error of no more than 3% stipulated by the power system, it is likely to be nonconvergence in the training, namely when the neural network training and learning samples at the same time, it is difficult to guarantee the global optimum, and it is easy to fall into local optimum loop, which may affects the precision of prediction.In order to continue to obtain the predict results of given data, we need to use other prediction methods.The solution given in this paper is to add judgment conditions in the neural network model, like training 100000 times.When the training sample data is not convergence, the data automatically fill to polynomial regression model, terminating the neural network prediction and jumping out of local minimum dead circulation.The polynomial regression model is established and prediction results are obtained.The combination method is trained and learned automatically intelligently.It has high accuracy and generalization ability, and can be

The input layer
The hidden layer The output layer general in the short-term economic indicators forecasting work.

Particle Swarm Optimization Algorithm
Particle swarm optimization (PSO) is a kind of evolutionary algorithm, derived from the observation of the birds' behavior of searching for food.The conversion process of the motion of whole flock from disorderly to orderly comes from information shared by each individual birds in the flock, so as to find food [9].

The current position
The optimal position of individual ant

The optimal position of ant colony
The next position The method used by PSO to solve optimization problems is to initialize a group of random particles and find the optimal solution through several iterations.In the process of iterations, each particle updates its direction and position constantly according to two extreme values.The first one is the optimal solution found by the particle itself, called individual extreme pbest , and the other one is the current optimal solu- tion found by the entire particle swarm, named global extreme gbest .At the beginning of the iterations, the position of each initialized particle is the individual extreme, while the best position of the particle swarm is the global extreme.After all of the particles in the swarm complete the first iteration, we should compare the position front and rear of each particle, and update the individual extreme with the optimal solution in this iteration if the new position is better than the previous one.Then, we need to get the optimal solution throughout individual extremes of all particles in the swarm as global extreme by comparison, and update the global extreme if the new one is better than the old one.The final global extreme obtained through these cycle iteration operations determines the optimal solution [12].
After obtaining individual extreme and global extreme in the process, each particle needs to update its velocity and position according to the following formulas: v i, j+1 = wv i, j + c 1 * random() *( pbest i, j !P i, j ) + c 2 * random()*(gbest i, j !P i, j ) (1) Where, v i, j denotes the velocity of the i particle after j it- erations, P i, j means the position of the i particle after j iter- ations, pbest i, j and gbest i, j are on behalf of the individual extreme and global extreme of the i particle after j iterations, w is the inertia weight of the updated speed to the speed of pre update, random() is a random number within 0,1 ( ) , 1 c and c 2 are learning factors within 0,2 ( ! " .The velocity of particles in the swarm is limited in 0,v max ( ) , and the updated value should be replaced with max v if it exceeds maximum max v in the process of iterations.
In the process of particle swarm optimization algorithm, particle shares global extreme value to other particles within the group.This one-way flow of shared information and data makes the whole search process follow the group within the current optimal solution.Therefore, the initial particle swarm optimization algorithm has fast global convergence capability.

Ant colony optimization algorithm
ACO is inspired by Italy scholar Dorigo M from the foraging behavior of real ant colony in nature.He found that an individual ant doesn't have much wisdom or master the nearby geographic information.But the colony can find an optimal path from nest to food sources.Dorigo.M and other researchers proposed ACO theory in 1991, attracting research enthusiasm of many scholars.The basic ACO model consists of the following three equations: Where, m is the number of ants, n is the number of iterations, i is the position of ants, j is the position where ants can reach, ! is the set of the position where ants can reach, ! ij is the heuristic information, which means the visibility of the path from i to j ,named , L k is the objective function, !ij is the pheromone intensity of the path from i to j , ! ij k is the number of pheromone left by ants on the path from i to j , ! is the weight of the path, ! is the weight of the heuristic information, ! is the evaporation factor of the number of pheromone on the path, Q is coefficient of the pheromone quality, and P ij k denotes the transition probabilities of the NO k ant moving from i to j .This paper proposes a hybrid algorithm based on particle swarm and ant colony.In the algorithm process, this paper uses the particle swarm optimization algorithm for fast global search to determine the parameters of ant colony and transform the better value into initial information pheromone.Then, using ant colony algorithm for path searching, we put the length of optimal solution, the running time and the numbers of iterations calculated by this set of parameters into the PSO algorithm and update the velocity and position of each particle according to formula until we get the optimal solution for ant colony algorithm [13,14].

EMPIRICAL ANALYSIS
Draw total output value of industry and agriculture and consumption statistics of each year in a period time in a given area.Gross output value of industry and agriculture are in constant prices in a year.
The paper chooses the training and prediction samples to establish the BP neural network, including 4 consecutive years of consumption data and fifth years of gross value of industrial output data as input and fifth years of consumption data as output.The neural network model contains only one hidden layer and the number of neurons in the hidden layer is selected by trial and error method.We first set fewer neurons, and gradually increase the number of neurons, until the network training error reaches the expected range.In order to avoid falling into local minimum for neural network, particle swarm and ant colony optimization are introduced to train the weights and threshold for the global optimum [13,14].
We sequentially select 4 continuous data as the first 4 input variables from the 37 consumption data.The fifth input variables are the gross value of industrial output data of fifth years.As shown in Table 1.The input data table is established as shown in Table 2. Table 1.Input values of the model.

Electricity
After normalization，the value of each variable is between 0,1 !" # $ ，which helps to eliminate the effects of dimen- sionless． After the completion of neural network training, we still use the input value of training data for simulation and get the forecasting data.The historical data and the change trend of the forecasting value are shown in Fig. (2).Overall, the neural network after optimization has good fitting degree.This paper choose Mean Absolute Error (MAE) to quantitative evaluation predict results [15]． The mean absolute error of each model is as shown in Table 3.In the error evaluations, the forecasting results of proposed model are the best fit．The error value of this model is 2.5 in MAE evaluation while it separately ups to 5.73， 12.87, 43.65 and 13.38 of the four methods for comparison.As a comparison, the neural network based on ACO and PSO forecasting model has obvious advantages.It has more accurate prediction to get better result and the model has high accuracy and generalization ability, which can be used to power load forecasting.

CONCLUSION
This paper presents an intelligent model of neural network based on the hybrid optimization of ACO and PSO.It is applied to short-term load forecasting and has extremely important practical significance for the power sector plan-

consumption the annual electricity consumption of 4 years before forecasting year the annual electricity consumption of 3
years before forecasting year the annual electricity consumption of 2 years before forecasting year the annual electricity consumption of 1 years before forecasting year Gross industrial production Gross industrial production of the forecasting year