A regional pretraining-classification-selection forecasting system for wind power point forecasting and interval forecasting

https://doi.org/10.1016/j.asoc.2021.107941Get rights and content

Highlights

  • A novel forecasting system is proposed to forecast wind power.

  • Multiple predictors are taken as candidate models for each prediction point.

  • Convolutional neural network is used as classifier to determine optimal predictors.

  • Reduction coefficients are optimized to adjust interval width.

Abstract

Wind power forecasting is extremely crucial for power market transactions and power system operation. Although a lot of researches have concentrated on wind power forecasting, the forecasting performances were confined and the forecasting models were only suitable for providing good performance at a few sites of investigation. To bridge these gaps, a novel regional pretraining-classification-selection wind power forecasting system is proposed in this paper based on four modules-pretraining module, classification module, point forecasting module, and interval forecasting module, which effectively improves forecasting performance and extends the applicability to different data characteristics. 10-min wind power data obtained from 20 datasets are used to verify the forecasting ability of the proposed forecasting system. The experimental analyses and discussions reveal that the proposed forecasting system is accurate and reliable for achieving high-quality wind power point and interval forecasting results. Thus, it could provide useful references for wind producers and managers in power system dispatch and operation.

Introduction

In recent decades, the concept of sustainable energy development has increased significantly worldwide, triggering huge reformation in energy structure [1]. Wind power, as an effective alternative of fossil fuels, occupies more and more shares in energy portfolios. It was reported that by the end of 2019, the global total installed wind power capacity was 650 GW, and in 2019, the new global wind power installed capacity amounted to 60.4 GW.1 Although wind power integration can bring attractive economic and environmental benefits, it is also accompanied by inevitable operational and planning risks because of the randomness and fluctuation of wind power [2], [3]. Thus, exploring accurate and stable wind power forecasting technologies is of great significant for decreasing operating cost and enhancing the reliability and feasibility of power generation system [4], [5].

Current wind power forecasting technologies are broadly divided into three categories, physical technologies [6], statistical technologies [7], and artificial intelligence (AI) technologies. Physical technologies mainly exploit measured physical information, such as temperature, humidity, and pressure, as input to forecast wind power [3]. Numerical weather prediction (NWP) plays a crucial role in physical technologies. Considering the high information collection cost and the intricate computation of physical technologies, huge challenges are posed to short-term wind power forecasting. Statistical technologies, as autoregressive moving average model (ARMA) [8], autoregressive integrated moving average (ARIMA) [9], establish the connection between the past and future wind power time series, which are more time-saving and easy to calculate. However, due to the precondition of linear patten, statistical technologies are not applicable to address nonlinear data patten. In this context, AI technologies have been developed to further depict the complex and nonlinear wind power data information. Common AI technologies include artificial neural networks (ANNs) [10] and support vector machine (SVM) [11]. However, AI technologies always suffer from the deficiency of local minimization and overfitting [12].

Based on above analysis, researchers have found that only single benchmark forecasting model cannot conduct effective wind power forecasting. Thus, current research direction about wind power forecasting has been converted to the development of hybrid models, which is constructed based on benchmark forecasting models, data preprocessing technologies, intelligent optimization algorithms, and other useful technologies [13]. Hybrid models can be concluded as two categories: hybrid models with single predictor, and hybrid models with multiple predictors. For the hybrid models with single predictor, only one predictor is employed to conduct the forecasting of wind power. For the hybrid models with multiple predictors, multiple predictors are employed to conduct the forecasting of wind power, and the forecasting results of these predictors are combined based on certain technology (also called combined models or ensemble models). Relevant literatures about hybrid models are listed in Table 1a, Table 1b.

The shortcomings of the above forecasting models can be summarized as follows:

(1) There are inevitable weaknesses in single forecasting models, thus, they cannot effectively perform renewable energy generation forecasting. Although hybrid models integrate single forecasting models with data preprocessing methods and intelligent optimization algorithms, they are also weak in the forecasting with different data characteristics under different situations.

(2) Hybrid models with multiple predictors can integrate the forecasting results of different forecasting models, however, these models still have some defects. Concretely, for combined models, the multicollinearity in the forecasting results of multiple predictors is a threatening issue, which will greatly impact forecasting performance. For ensemble models, the multicollinearity problem can be effectively avoided, however, only a single dataset can be forecasted at a time, occupying a lot of operation resources.

To bridge these shortcomings, a regional pretraining- classification-selection forecasting system (RPCSFS) is proposed to forecast wind power in different datasets and help decision-makers make effective decisions. To be specific, the pretraining for multiple candidate predictors is conducted based on training data. Then, all candidate predictors are used to forecast data in validation set, from which the optimal forecasting model category of each point in every dataset is determined. Moreover, the data in validation set and the above selected optimal forecasting model category are taken as the input and output of the CNN. The aim of the process is to better train CNN. CNN is selected as the classifier because it possesses great self-learning and feature extraction ability, which is conducive to capture the nonlinear feature of original wind power series [23]. Moreover, compared with other classifiers (i.e., LSTM, SVM, and ELM) in experiment, CNN can provide high classification accuracy. Next, in point forecasting (PF) process, the first six-period test data are firstly taken as the input of CNN, and the output is the optimal forecasting model category of the forecasting point. In this way, the final optimal PF results can be obtained through imputing the first six-period test data into pretrained optimal forecasting model. Furthermore, to provide more valuable reference information for decision-makers, interval forecasting (IF) for wind power is further conducted. Several common statistical distributions are used to measure the uncertainty characteristics of the forecasting error of the verification set, and we can further infer the probability distribution of the forecasting error of the test set. Finally, a modified multi-objective dragonfly algorithm (MMODA) is used to optimize the reduction coefficient of the forecasting interval width. Relative to traditional single-objective optimization algorithms with only one optimization objective, the adopted multi-objective algorithm considers both IF reliability and accuracy, which improves the effectiveness and reliability of forecasting interval [31]. Besides, considering that most optimization algorithms are obsessed with the issues of local optimum and slow convergence speed in the late stage of the algorithm, two valid strategies (i.e., the elite opposition learning (EOPL) scheme and the exponential function steps-based (EFS) scheme) are introduced into the modified optimization algorithm [32]. EOPL can solve the issue of local optimum while EFS can address the issue of slow convergence speed. Both theoretical proof and simulation experiments reveal that the proposed RPCSFS with optimized reduction coefficient can obtain the most accurate and reliable IF results.

The main contributions of this paper are as follows:

(1) A RPCSFS is proposed to forecast the wind power of multiple stations at one time, which saves computing resources, improves forecasting efficiency, and effectively integrates the optimal forecasting results of multiple predictors. The proposed forecasting system overcomes the multicollinearity issues of combined models and the low prediction efficiency of ensemble models, and obtains effective forecasting performance.

(2) In PF process, CNN is used as classifier to determine the optimal forecasting model category. To prompt the classification performance of CNN, data in validation set are used to train CNN, and then, the trained CNN can accurately recognize the category of the optimal forecasting models at each data point in the test set.

(3) In IF process, a reduction coefficient optimized by a modified multi-objective optimization algorithm is proposed to adjust the forecasting interval width. Theoretical proof verifies the effectiveness of the optimized reduction coefficient, and the simulation experiments indicate that the proposed RPCSFS obtain reliable and accurate IF results with high coverage rate and narrow interval width.

The rest of this paper is organized as follows: Section 2 introduces the methods and the frame of the proposed RPCSFS. Section 3 shows the experiments and corresponding analysis. The deeper discussions are conducted in Section 4. Section 5 is the conclusion and the future direction.

Section snippets

Design of the proposed forecasting system

In this section, the involved significant methods used to construct the proposed RPCSFS, including CNN, MMODA, and the frame of the proposed forecasting system, are introduced in detail.

Experiment and analysis

In this section, the studied data and involved evaluation criteria are introduced. Subsequently, three experiments are conducted to verify the classification performance of CNN and the PF and IF performance of the proposed RPCSFS.

Discussion

Three discussions are performed in this section, including: comprehensive performance evaluation, improvement percentage of the proposed forecasting system, and comparative analysis with other existing methods.

Conclusion

Accurate wind power forecasting is beneficial to maintain secure operation, curtail ancillary service costs, and optimize management of the power system, which has drawn great attention by researchers in recent years. Due to various data characteristic in different datasets, previous studies have been limited by providing high-quality forecasting results for more than one dataset, simultaneously. To bridge this gap, in this study, a novel wind power forecasting system based on the CNN

Limitation and future direction

The limitations of the proposed forecasting system and the future directions lie in the following three aspects: (1) The run time of CNN is relatively long and multiple predictors require parallel computation, which means high computational cost are required. In the future, more measures should be applied to accelerate its computation. (2) Only single-step forecasting is conducted in this paper, and in the future, multi-step forecasting should be taken into consideration. (3) Only original wind

CRediT authorship contribution statement

Jianzhou Wang: Writing – review & editing. Lifang Zhang: Writing – original draft. Chen Wang: Software, Validation. Zhenkun Liu: Visualization, Supervision.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

This manuscript is supported by the Department of Education of Liaoning Province of China (No. LN2019Z13).

References (50)

  • WuZ. et al.

    Combined model with secondary decomposition-model selection and sample selection for multi-step wind power forecasting

    Appl. Energy

    (2020)
  • LiuH. et al.

    Corrected multi-resolution ensemble model for wind power forecasting with real-time decomposition and Bivariate Kernel density estimation

    Energy Convers. Manag.

    (2020)
  • ChenC. et al.

    Medium-term wind power forecasting based on multi-resolution multi-learner ensemble and adaptive model selection

    Energy Convers. Manag.

    (2020)
  • DingM. et al.

    A gated recurrent unit neural networks based wind speed error correction model for short-term wind power forecasting

    Neurocomputing

    (2019)
  • WangG. et al.

    A hybrid wind power forecasting approach based on Bayesian model averaging and ensemble learning

    Renew. Energy.

    (2020)
  • JingG. et al.

    An energy-saving control strategy for multi-zone demand controlled ventilation system with data-driven model and air balancing control

    Energy

    (2020)
  • KisvariA. et al.

    Wind power forecasting – A data-driven method along with gated recurrent neural network

    Renew. Energy.

    (2021)
  • HeY. et al.

    Uncertainty analysis of wind power probability density forecasting based on cubic spline interpolation and support vector quantile regression

    Neurocomputing

    (2021)
  • González-SopeñaJ.M. et al.

    An overview of performance evaluation metrics for short-term statistical wind power forecasting

    Renew. Sustain. Energy Rev.

    (2021)
  • YangM. et al.

    Day-ahead wind power forecasting based on the clustering of equivalent power curves

    Energy

    (2021)
  • NaikJ. et al.

    Short-term wind speed and wind power prediction using hybrid empirical mode decomposition and kernel ridge regression

    Appl. Soft Comput. J.

    (2018)
  • HeY. et al.

    Short-term wind power prediction based on EEMD–LASSO–QRNN model

    Appl. Soft Comput.

    (2021)
  • ZhouM. et al.

    Multi-objective prediction intervals for wind power forecast based on deep neural networks

    Inf. Sci. (Ny)

    (2021)
  • BikcoraC. et al.

    Density forecasting of daily electricity demand with ARMA-GARCH, CAViaR, and CARE econometric models

    Sustain. Energy Grids Netw.

    (2018)
  • NiuX. et al.

    A combined model based on data preprocessing strategy and multi-objective optimization algorithm for short-term wind speed forecasting

    Appl. Energy

    (2019)
  • Cited by (0)

    View full text