A new approach to very short term wind speed prediction using k-nearest neighbor classification

https://doi.org/10.1016/j.enconman.2013.01.033Get rights and content

Abstract

Wind energy is an inexhaustible energy source and wind power production has been growing rapidly in recent years. However, wind power has a non-schedulable nature due to wind speed variations. Hence, wind speed prediction is an indispensable requirement for power system operators. This paper predicts wind speed parameter in an n-tupled inputs using k-nearest neighbor (k-NN) classification and analyzes the effects of input parameters, nearest neighbors and distance metrics on wind speed prediction. The k-NN classification model was developed using the object oriented programming techniques and includes Manhattan and Minkowski distance metrics except from Euclidean distance metric on the contrary of literature. The k-NN classification model which uses wind direction, air temperature, atmospheric pressure and relative humidity parameters in a 4-tupled space achieved the best wind speed prediction for k = 5 in the Manhattan distance metric. Differently, the k-NN classification model which uses wind direction, air temperature and atmospheric pressure parameters in a 3-tupled inputs gave the worst wind speed prediction for k = 1 in the Minkowski distance metric.

Highlights

► Wind speed parameter was predicted in an n-tupled inputs using k-NN classification. ► The effects of input parameters, nearest neighbors and distance metrics were analyzed. ► Many useful and reasonable inferences were uncovered using the developed model.

Introduction

Wind energy has a critical role in the electricity generation based on renewable energy and its utilization ratio is increasing remarkably in the world. The most significant indicator of this case is that the global cumulative installed wind power capacity reached to 198 GW in 2010, while it was only 6.1 GW in 1996 [1], [2]. In wind energy systems, wind power generated by wind turbines has an intermittent and variable structure due to the stochastic nature of wind speed [3]. So, dispatchers need to predict the wind power for wind turbine control, pre-load sharing, power system management, energy trading, maintenance and repair of wind turbines [4]. As a result, wind speed is a crucial parameter for wind power integration.

Many methods have been developed in the field of wind speed prediction. These methods are classified into four categories as physical models, conventional statistical models, spatial correlation models and artificial intelligence models [5].

Physical models use physical data such as temperature, pressure, orography, roughness and obstacles in order to predict the wind speed, but good at long term wind speed prediction [5], [6]. Some of these models are Weather Research and Forecast (WRF) model [7], High Resolution Model (HRM) [8], Consortium for Small Scale Modeling (COSMO) [9], Mesoscale Model 5 (MM5) [10] and ETA model [11]. The spatial correlation methods consider the wind speed time series of the predicted sites and their neighboring sites in order to predict the wind speed, but these models have some challenges such as the need for measuring and transmitting the wind speeds of many correlated sites [5]. Fuzzy logic, neural network and fuzzy neural network models based on spatial correlation were implemented in [12], [13], [14].

Conventional statistical models use a mathematical model of the problem and give better results than physical models in very short term wind speed prediction [5], [6]. Some of these models are autoregressive (AR) models [15], autoregressive moving average (ARMA) models [16], [17], autoregressive integrated moving average (ARIMA) models [18], [19], Markov chain model [20] and Kalman filtering [21]. Artificial intelligence methods have a wide range of application in the field of wind speed prediction [5]. Radial basis functions (RBFs) [22], recurrent neural networks (RNNs) [23], multi-layer perceptrons (MLPs) [24], support vector machines (SVMs) [25], Bayesian learning model [26], fuzzy logic model [27], adaptive neuro-fuzzy inference model [28], [29] and k-means clustering model [30] were implemented in the literature.

The aim of this paper is, on the one hand, to predict the wind speed parameter using wind direction, air temperature, atmospheric pressure and relative humidity parameters in an n-tupled inputs with the k-NN classification model. On the other hand, the number of nearest neighbors, the dimension of input parameters and the selected distance metric were analyzed for investigating their effects on the prediction performance. Many useful and reasonable inferences were uncovered for wind speed prediction in this paper.

The rest of the paper is organized as follows. The developed k-NN classification model and the selected distance metrics are introduced in Section 2. Section 3 describes the dataset used for this study and explains the data cleaning process. The wind speed prediction results based on 3-tupled and 4-tupled inputs are presented and discussed in Section 4. Section 5 concludes the work.

Section snippets

k-Nearest neighbor classification

Data mining is a process of knowledge extraction used for revealing previously unknown, hidden, meaningful and useful patterns in databases [31]. k-NN classification has a wide range of application in the field of data mining and based on using specific training instances to make a class prediction for a new unclassified instance [32]. Euclidean distance metric is often used for measuring the closeness of the data points in k-NN classification [33]. Differently, Manhattan and Minkowski distance

Data description and data preprocessing

The dataset used in this study was generated by a meteorological station placed in Poyracık, Turkey. Poyracık has an elevation of 125 m and its latitude and longitude are 39°9′N and 27°29′E, respectively. Its annual average temperature, relative humidity and rainfall are 17.9 °C, 54.52% and 57.05 kg/m2, respectively [37]. In addition, its wind climate regime depends on Etesian winds in summer as a result of the high pressure over the Balkans and the heating of the Anatolian plateau [38]. The total

Multi-tupled inputs in wind speed prediction

A small error in wind speed prediction leads to a large error in wind power prediction [42]. Therefore, it is needed to predict the wind speed parameter in an accurate way. For this purpose, a k-NN classification model was developed and used for predicting the wind speed parameter in this study. Unlike the studies in the literature, the model developed predicts the wind speed parameter in an n-tupled inputs. In here, each dimension represents each input parameter as described in [43], [44]. For

Conclusions

This paper presented a new approach based on k-NN classification to short term wind speed prediction using multi-tupled inputs. The k-NN classification model developed included not only Euclidean, Manhattan and Minkowski distance metrics but also MAE, MAPE and NRMSE metrics. The numerical classification results were achieved rapidly by means of the object oriented programming techniques adapted to the model developed. Besides, the observed and the predicted wind speeds along with their error

References (48)

  • R.G. Kavasseri et al.

    Day-ahead wind speed forecasting using f-ARIMA models

    Renew Energy

    (2009)
  • A. Shamshad et al.

    First and second order Markov chain models for synthetic generation of wind speed time series

    Energy

    (2005)
  • F. Cassola et al.

    Wind speed and wind energy forecast through Kalman filtering of numerical weather prediction model output

    Appl Energy

    (2012)
  • S.L. Goh et al.

    Complex-valued forecasting of wind profile

    Renew Energy

    (2006)
  • P. Lopez et al.

    Effect of direction on wind speed estimation in complex terrain using neural networks

    Renew Energy

    (2008)
  • M.A. Mohandes et al.

    Support vector machines for wind speed prediction

    Renew Energy

    (2004)
  • G. Li et al.

    Bayesian adaptive combination of short-term wind speed forecasts from neural network models

    Renew Energy

    (2011)
  • M. Mohandes et al.

    Estimation of wind speed profile using adaptive neuro-fuzzy inference system (ANFIS)

    Appl Energy

    (2011)
  • R. Ata et al.

    An adaptive neuro-fuzzy inference system approach for prediction of tip speed ratio in wind turbines

    Expert Syst Appl

    (2010)
  • A. Kusiak et al.

    Short-term prediction of wind power with a clustering approach

    Renew Energy

    (2010)
  • S.L. France et al.

    Distance metrics for high dimensional nearest neighborhood recovery: compression and normalization

    Inf Sci

    (2012)
  • C. Chourpouliadis et al.

    Comparative study of the power production and noise emissions impact from two wind farms

    Energy Convers Manage

    (2012)
  • Y. Liu et al.

    A fast differential evolution algorithm using k-nearest neighbour predictor

    Expert Syst Appl

    (2011)
  • X. Zhao et al.

    Review of evaluation criteria and main methods of wind power forecasting

    Energy Procedia

    (2011)
  • Cited by (104)

    View all citing articles on Scopus
    View full text