Short-Term Load Forecasting with Improved CEEMDAN and GWO-Based Multiple Kernel ELM

Taiyong Li,Zijie Qian,Ting He

doi:10.1155/2020/1209547

Abstract

Short-term load forecasting (STLF) is an essential and challenging task for power- or energy-providing companies. Recent research has demonstrated that a framework called “decomposition and ensemble” is very powerful for energy forecasting. To improve the effectiveness of STLF, this paper proposes a novel approach integrating the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), grey wolf optimization (GWO), and multiple kernel extreme learning machine (MKELM), namely, ICEEMDAN-GWO-MKELM, for STLF, following this framework. The proposed ICEEMDAN-GWO-MKELM consists of three stages. First, the complex raw load data are decomposed into a couple of relatively simple components by ICEEMDAN. Second, MKELM is used to forecast each decomposed component individually. Specifically, we use GWO to optimize both the weight and the parameters of every single kernel in extreme learning machine to improve the forecasting ability. Finally, the results of all the components are aggregated as the final forecasting result. The extensive experiments reveal that the ICEEMDAN-GWO-MKELM can outperform several state-of-the-art forecasting approaches in terms of some evaluation criteria, showing that the ICEEMDAN-GWO-MKELM is very effective for STLF.

Highlights

Accurate load forecasting plays a significant role for the participants in the electricity industry because it can provide a safe and reliable automatic management for a smart grid
We conduct extensive experiments to compare the proposed ICEEMDAN-grey wolf optimization (GWO)-multiple kernel extreme learning machine (MKELM) with many state-of-the-art approaches, and the results demonstrate that the ICEEMDAN-GWO-MKELM is able to outperform the compared approaches in most cases
When we use multiple kernel Extreme Learning Machine (ELM) optimized with GWO (GWOMKELM) for Short-term load forecasting (STLF), it achieves the best results in 29 out of 60 cases, which indicates that GWO-MKELM outperforms other single models significantly. e possible reason is that GWO-MKELM can optimize the weights and parameters of the multiple kernels for ELM to improve the representation ability of the kernels

Summary

Introduction

Accurate load forecasting plays a significant role for the participants in the electricity industry because it can provide a safe and reliable automatic management for a smart grid. Time series-based models, such as random walk, moving average (MA), autoregressive integrated MA (ARIMA), ARIMA with explanatory variable (ARIMAX), and generalized autoregressive conditional heteroskedasticity (GARCH), are widely used in load forecasting [3, 4]. Cui and Peng introduced temperature into ARIMA to propose an improved ARIMAX to deal with the mutation data structures [4]. Because these time series-based models are usually built on the assumption that the load data are with the characteristics of linearity and stationarity, which are not always met in practical load data, the forecasting accuracy is limited. Erefore, it is necessary to use other models instead of time series-based models to improve load forecasting accuracy. Typical AI models include support-vector regression (SVR) and its extension least-squares SVR (LSSVR) [5, 6], artificial neural network (ANN) [7,8,9], extreme learning machine (ELM) [10], sparse Bayesian learning (SBL) [11, 12], deep learning [13]

Results

Discussion

Conclusion