Short Term Wind Speed Prediction Using a New Hybrid Model With Passive Congregation

Abstract

Short term wind speed predicting is essential in using wind energy as an alternative source of electrical power generation, thus the improvement of wind speed prediction accuracy becomes an important issue. Although many prediction models have been developed during the last few years, they suffer a poor performance becausetheir dependency on performing only the local search without the capability in performing the global search in the whole search space.To overcome this problem, we propose a new passive congregation term to the standard hybrid Genetic Algorithm / Particle Swarm Optimization (GA/ PSO) model in training Neural Network (NN) wind speed predictor. This term is based on the mutual cooperation between different particles in determining new positions rather than their selfish thinking. Experiment study shows significantly the influence of the passive congregation term in improving the performance accuracy compared to the standard model. General TermsWind Speed Prediction, Key wordsParticle Swarm Optimization, Genetic Algorithm, Neural Networks, Passive Congregation. INTRODUCTIONWind energy is one of the lowest costs of electricity production among renewable energy sources, but is visible only as long as weather conditions allow, so to maintain economical dispatch of wind generation electricity, it is important to make short term predictions of future wind speed which directly affects generation capacity [1]. Without this ability, a wind farm operator is prone to allocate more generation units or supplemental energy reserves than necessary in order to ensure budgeted electricity outputs are met, with an end result of increased operating costs [2]. Thus, the further prediction accuracy improvement of the wind speed predication becomes a fundamental issue in the wind industry [2-3].In short term wind speed prediction, Feedforward Neural Networks (FNN) trained using tabu search [4] and Recurrent Neural Networks (RNN) [5] have been used.Particle Swarm Optimization (PSO) which is a population based stochastic algorithm was used in training both FNN and RNN wind speed predictors to enhance their prediction accuracy [2]. Also, Genetic Algorithm (GA) operators was introduced to increase the training accuracy of PSO learning algorithm, to form GA/ PSO hybrid model to enable searching new regions in the search space [6][7].However, in GA/ PSO model, the PSO algorithm is restricted in doing the local search around its best position and the global best position without any cooperation with other particles in the swarm while the GA operators are responsible in doing the global search by making particles fly to new regions in the search space leading to poor model performance in terms of accuracy, convergence speed and robustness [8].To overcome this problem and improve the model performance, each particle in the swarm should be capable of performing the global search and not to be only restricted in local search [9]. Thus, we introduce our new proposed hybrid (GA/PSO) model with passive congregation inspired ...

Increasing Wind Speed Prediction Accuracy Using Hybrid Model with Passive Aggregation

“…The hybrid model consists of three parts, first the Particle Swarm Optimization (PSO), second the passive congregation term and the last part is the Genetic Algorithm (GA) operators [11].…”

“…Passive congregation is an attraction from one individual to others but does not display social behaviors and any individuals may have law fidelity to other group members if they have little or no genetic relation to each other's [11], [13]. To fully represent this mechanism, other particles positions should be included by replacing the cognitive term with the passive congregation term PC i k + 1 and can be expressed by the following equation [11].…”

“…where β 1 and β 2 are two uniformly distributed random variables in the range [0, 1]. For the mutation process, all variables in the new created child particles are equally probable to be mutated [11].…”

“…Thus the need for accurate prediction model increases to handle these fluctuation movements. The number of generation is selected to be 150 with swarm size of 50 particles, all constants are set to be one and a breading ratio is 0.1 [11].The hourly wind speed data is divided into two parts the first 80% of the data will be taken as a training data while the remaining 20% will be the testing data and the input to the SDRNN predictor are the last four hours wind speed and the previous predictor output [11]. The actual and predicted wind speed using our proposed passive aggregation model for the last three days of the month is shown in Fig.3and 4.…”

“…However, both GAs and PSO has poor prediction accuracy due to the premature convergence to local minimum [10]. To overcome this problem, passive congregation term was combined with both PSO and GAs operators to form a hybrid model that increases prediction accuracy by avoiding premature convergence and was applied in short term wind speed prediction in Canada [11]. In this model, Sigmoid Diagonal Recurrent Neural Network (SDRNN) was used as wind speed predictor [11] because it was proven that it t is better than both FNN and RNN in practical dynamical nonlinear applications [12].…”