A new short-term load forecast method based on neuro-evolutionary algorithm and chaotic feature selection Sajjad Kouhi a,b,⇑ , Farshid Keynia c , Sajad Najafi Ravadanegh b a Heris Branch, Islamic Azad University, Heris, Iran b Smart Distribution Grid Research Lab, Electrical Engineering Department, Azarbaijan Shahid Madani University, Tabriz, Iran c Energy Department, Graduate University of Advanced Technology, Kerman, Iran article info Article history: Received 18 May 2013 Received in revised form 27 March 2014 Accepted 12 May 2014 Available online 19 June 2014 Keywords: Short-term load forecast Neural network Chaotic time series Feature selection Reconstructed phase space Differential Evolutionary abstract In competitive environment of deregulated electricity market, short-term load forecasting (STLF) is a major discussion for efficient operation of power systems. Therefore, the area of electricity load forecast- ing is still essential need for more accurate and stable load forecast algorithm. However, the electricity load is a non-linear signal with high degree of volatility. In this paper, a new forecasted method based on neural network (NN) and chaotic intelligent feature selection is presented. The proposed feature selec- tion method selects the best set of candidate input which is used as input data for the forecasted. The theory of phase space reconstruction under Taken’s embedding theorem is used to prepare candidate fea- tures. Then, candidate inputs relevance to target value are measured by using correlation analysis. Fore- cast engine is a multilayer perception layer (MLP) NN with hybrid Levenberg–Marquardt (LM) and Differential Evolutionary (DE) learning algorithm. The proposed STLF is tested on PJM and New England electricity markets and compared with some of recent STLF techniques. Ó 2014 Elsevier Ltd. All rights reserved. Introduction Short-term load forecasting (STLF) plays a key role in operation of both traditional and deregulated power systems. In deregulated electricity market, STLF is a useful tool for economic and reliable operation of power system. Many operating decisions are based on load forecast such as: dispatch scheduling of generating produc- tion, reliability and security analysis and maintenance plan for generators [1]. Therefore, load forecasts are vital for the market players in competitive electricity market [2]. Hence, improving the accuracy of STLF can increase the appropriateness of planning and scheduling and reduce operational costs of power systems. Load forecasting algorithms are includes traditional methods and modern intelligent methods [3]. Traditional methods based on mathematical statistics including regression analysis method [4], Kalman filtering method [5], autoregressive integrated moving average (ARIMA) [6], Box–Jenkins models [7], state space model [8], exponential soothing [9], and etc. These methods have the advantage of mature technology and simple algorithm, but these are based on linear analysis and none of them can forecast the non-linear load series accurately [3]. The modern intelligent forecasting methods have shown better performance for non-linearity of the time series. Also, they do not require any complex mathematical formulations or quantitative correlation between inputs and outputs. Effective utilization of intelligent algorithms in the context of ill-defined processes (such as load time series), have led to their wide application in STLF [10,11]. The intelligent algorithms including artificial neural net- work (ANN) based methods [11–13]. It seems that the ANN learned well the training data, but it may encounter great forecast error in the test phase. The area of the electricity load forecasting is still essential need for more accurate load forecast. Especially, there is a requirement for efficient feature selection and input/output map- ping algorithm. Knowledge based expert system (KBES) approach [14,15], extracts rules from received relevant information. How- ever, the training procedure of a KBES model is a time-consuming procedure. Electricity load is a time variant, non-linear and volatile signal. Design of the input vector to the forecast engine is an important pre-processing phase. It plays an important role in forecast accu- racy. There are two general approaches for designing of the input vector: wrapper and filter methods [16]. Feature selection is wrapped around a learning algorithm in the wrapper methods. In this method, the usefulness of a feature is judged by the estimated accuracy of the learning algorithm directly. For the electricity load forecasting with a large number of features, these methods are http://dx.doi.org/10.1016/j.ijepes.2014.05.036 0142-0615/Ó 2014 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. Tel.: +98 914 1284684; fax: +98 342 6233176. E-mail address: sajjadkouhi@gmail.com (S. Kouhi). Electrical Power and Energy Systems 62 (2014) 862–867 Contents lists available at ScienceDirect Electrical Power and Energy Systems journal homepage: www.elsevier.com/locate/ijepes