A deep learning based hybrid method for hourly solar radiation forecasting

https://doi.org/10.1016/j.eswa.2021.114941Get rights and content

Highlights

Abstract

Solar radiation forecasting is a key technology to improve the control and scheduling performance of photovoltaic power plants. In this paper, a deep learning based hybrid method for 1-hour ahead Global Horizontal Irradiance (GHI) forecasting is proposed. Specifically, a deep learning based clustering method, deep time-series clustering, is adopted to group the GHI time series data into multiple clusters to better identify its irregular patterns and thus providing a better clustering performance. Then, the Feature Attention Deep Forecasting (FADF) deep neural network is built for each cluster to generate the GHI forecasts. The developed FADF dynamically allocates different importance to different features and utilizes the weighted features to forecast the next hour GHI. The solar forecasting performance of the proposed method is evaluated with the National Solar Radiation Database. Simulation results show that the proposed method yields the most accurate solar forecasting among the smart persistence and state-of-the-art models. The proposed method reduces the root mean square error as compared to the smart persistence by 11.88% and 12.65% for the Itupiranga and Ocala dataset, respectively.

Introduction

Smart grids and cities concern with anticipating situations, with an efficient forecast of weather conditions (e.g., solar energy forecasting) and the possibility of making decisions in almost real-time. Solar forecasting is crucial in managing power network operations and solar photovoltaic applications (Huang et al., 2018, Lai et al., 2017, Wang et al., 2018, Wang et al., 2019). High penetration of intermittent renewable sources and the lack of anticipatory capabilities will result in uneconomic choices (such as the severe curtailment of generation) or lack of resilience when facing faults and disturbances. The accurate forecasting of Global Horizontal Irradiance (GHI) is beneficial to quality future power productions.

In previous studies, many data-driven approaches are proposed for short-term solar forecasting. These approaches can be mainly divided into three parts: the physical methods, the statistical methods, and the machine learning methods. The physical methods consist of a set of mathematical equations describing the physical state and dynamic motion of the atmosphere (Zhang et al., 2018), and its forecasting performance is highly affected by the sharp changes in meteorological variables. The statistical methods utilizing statistical analysis of the different input features for solar forecasting include the auto-regressive integrated moving average, the exponential smoothing, and the Markov Chain model (Shakya et al., 2017), among others. Recently, machine learning based approaches have been widely used in modeling, design and prediction in solar energy systems. The combination of two or more machine learning methods is also used to provide a more accurate solar forecasting result known as the hybrid model.

From the literature, a collection of classical machine learning methods has been proposed and applied in the solar energy system. For example, an Artificial Neural Network (Fermín et al., 2018) model was developed to predict the levelized electricity cost of two parabolic trough solar thermal power plants coupled with a fuel backup system and thermal energy storage (Boukelia et al., 2017). The Support Vector Machine was adopted (Ma et al., 2017) to upgrade the estimation accuracy of solar irradiance levels from photovoltaic electrical parameters. Also, the potential of the Random Forest method for estimating solar radiation using air pollution index was assessed (Sun et al., 2016).

With the development and improvement of deep learning algorithms which are one of the advanced machine learning methods, there are additional deeper learning methods being applied to renewable energy challenges. Convolutional Neural Network (CNN) was used (Sun et al., 2018, Sun et al., 2018) to forecast the solar Photovoltaic (PV) output using the contemporaneous images of the sky. The meteorological features were utilized (Qing and Niu, 2018) as the input for a Long Short-Term Memory neural network (LSTM) for day ahead hourly solar radiance prediction. Furthermore, studies (Ghimire et al., 2019, Yan et al., 2020) applied CNN to robustly extract features from predictive variables while the LSTM or GRU was utilized to absorb the features for solar radiation forecasting. The study (Feng and Zhang, 2020) adopted the CNN to forecast the solar PV output using the contemporaneous images of the sky, while the study (Zhen et al., 2020 firstly assigns the sky image to the corresponding class using the deep clustering method and then utilizes a corresponding hybrid deep learning method for PV power forecasting.

Due to the limitation with stand-alone machine learning methods, the hybrid model combining multiple machine learning methods was conducted to improve the solar forecasting accuracy. The Adaptive Neuro-Fuzzy Interface Systems and the Wavelet Neural Network are among the early generation of hybrid models (Faizollahzadeh et al., 2018, Fotovatikhah et al., 2018). The study (David et al., 2016) evaluated performances of an ensemble of Autoregressive Moving Average (ARMA) and Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models to establish solar irradiance probabilistic forecasts. Still based on the ensemble learning, two advance base models, namely extreme gradient boosting forest and deep neural networks (XGBDNN), were proposed for hourly Global Horizontal Irradiance forecast (Kumari & Toshniwal, 2021). A two-stage method Coral Reefs Optimization Extreme Learning Machine (CRO-ELM) was applied (Salcedo-Sanz et al., 2018a) to select useful features and use selected features for solar forecasting. Since solar features are highly influenced by weather conditions, the combination of clustering and regression was conducted. A novel clustering method TB_K-means was proposed (Azimi et al., 2016) to partition the solar data into several clusters where the multiple layer perceptron was developed for each cluster to forecast hourly solar radiation. The similar hybrid forecasting method can also be found in other studies (Li et al., 2017, Feng et al., 2018, Fu et al., 2019), where the traditional clustering methods were utilized to categorize the data based on the original solar radiation sequence or the pre-extracted features. The hierarchical clustering technique was utilized in a different way (Sun et al., 2018, Sun et al., 2018) where the K-means clustering was used to cluster the forecasting results of each sub-component generated from the Ensemble Empirical Mode Decomposition method, and a least square support vector regression was applied to ensemble the sub-component forecasts of each cluster. Similar study (Theocharides et al., 2020) was conducted where the K-means was applied to categorize the forecasted daily GHI and for each cluster, coefficients were obtained by linear regression to correct the forecasted outputs of the machine learning model.

Most of the existing hybrid methods utilizing the clustering techniques for solar forecasting usually adopt traditional clustering methods (Azimi et al., 2016, Li et al., 2017, Feng et al., 2018, Fu et al., 2019) which may result in sub-optimal clustering outcomes because the feature extraction and clustering are conducted in two separate independent stages rather than jointly considered. Besides, the GHI features (e.g., historical GHI, clear-sky GHI) and the meteorological features (e.g., temperature, wind speed) are often used to forecast the future solar irradiance (Jeon and Kim, 2020, Pan et al., 2020, Wu et al., 2020). The features are treated equally to forecast solar radiation in some previous studies. Intuitively, the historical solar irradiance should play a more important role than other features for solar forecasting. Some studies consider possible redundant features that may hurt the solar forecasting performance and thus utilize feature selection methods to select an optimal feature subset from the original feature set for improving solar forecasting performance (Salcedo-Sanz et al., 2018b, Niu et al., 2020, Qadir et al., 2021). In this work, instead of utilizing the feature selection, a deep learning based feature weighting method is proposed alternatively to automatically enhance more useful features and restrain less important features for solar forecasting. Because the feature weighting can be observed as a generalization of feature selection where the feature weights are not limited to 0 or 1.

The proposed deep learning based hybrid method in this work consists of the Deep Time-series Clustering (DTC) and the Feature Attention based Deep Forecasting (FADF). DTC groups the solar time series with similar patterns into the same clusters using high-level useful features extracted by the Gated Recurrent Unit neural network (GRU) (Chung et al., 2014), where the feature learning and clustering are learned jointly. Each cluster has a corresponding hourly solar predictor (i.e., the FADF) trained with the data in the corresponding cluster. The FADF of each cluster utilizes a Feature Attention Sub-network to determine the feature importance (Song et al., 2018) and sends the weighted input to the main GRU network for hourly solar forecasting. The major contributions of this work are as follows:

  • 1)

    A deep learning based clustering method (i.e., Deep Time-series Clustering, DTC) is proposed to group the solar irradiance (i.e., GHI) data with similar patterns into the same clusters. It optimizes the feature learning and clustering assignment simultaneously. DTC treats the feature learning and clustering assignment in two separate stages.

  • 2)

    A Feature Attention based Deep Forecasting method (FADF) is proposed for hourly solar radiation forecasting of each cluster grouped by the DTC. The FADF utilizes a feature attention mechanism to dynamically allocate different importance to different features at each time step for 1-hour ahead GHI forecasting.

  • 3)

    Extensive simulations are carried out to confirm the superiority of the proposed method. Simulation results on the National Solar Radiation Database show that the proposed method outperforms existing methods for solar forecasting in most cases.

The remainder of this paper is organized as follows. Section 2 describes the proposed method for 1-hour ahead GHI forecasting. Section 3 gives the simulation results and detailed discussion. Finally, a conclusion is given in Section 4.

Section snippets

The Hybrid method for one-hour ahead GHI forecasting

The training and testing phase of the proposed deep learning based hybrid method for 1-hour ahead GHI forecasting is shown in Fig. 1. The historical GHI time series of finite length (specified by the window size) is sent to the DTC to get its clustering label. Then, the FADF of the corresponding cluster 1) assigns the feature importance of the GHI features (including historical GHI, clear-sky GHI, clear-sky index and solar zenith angle) and the meteorological features (temperature, relative

Simulation setup

  • 1)

    Data

This study employs two 12-year (from 2005 to 2017) hourly datasets collected from the National Solar Radiation Database (Sengupta et al., 2018) to train and test the 1-hour ahead GHI forecasting models. One dataset is based on Itupiranga (latitude = 5.15° S, longitude = 49.34° W), Brazil and the other one is based on Ocala (latitude = 29.17° N, longitude = 82.14° W), Marion, Florida, United States.

In this work, all the hourly data from 2005 to 2014 were used as the training set, the data

Conclusions

In this paper, we propose a deep learning based hybrid method for 1-hour ahead GHI forecasting. The proposed method adopts the Deep Time-series Clustering (DTC) to group the GHI time series data into multiple clusters to better identify its irregular patterns and thus to provide a better clustering performance. Then, the Feature Attention Deep Forecasting (FADF) model which is capable of dynamically weighting the input features and using the weighted features to forecast the next hour GHI is

CRediT authorship contribution statement

Chun Sing Lai: Conceptualization, Methodology, Validation, Writing - original draft, Writing - review & editing, Resources, Supervision. Cankun Zhong: Conceptualization, Methodology, Software, Investigation, Data curation, Formal analysis, Writing - original draft, Writing - review & editing. Keda Pan: Writing - review & editing, Data curation, Validation, Resources. Wing W.Y. Ng: Conceptualization, Methodology, Writing - review & editing, Resources, Project administration, Funding acquisition,

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.

Acknowledgements

This work is supported by National Natural Science Foundation of China under Grants 61876066 and 61572201, Guangzhou Science and Technology Plan Project 201804010245, Department of Finance and Education of Guangdong Province 2016 [202]: Key Discipline Construction Program, China; the Education Department of Guangdong Province: New and Integrated Energy System Theory and Technology Research Group [Project Number 2016KCXTD022]; Brunel University London BRIEF Funding, UK.

References (43)

  • M. Sengupta et al.

    The national solar radiation database (NSRDB)

    Renewable and Sustainable Energy Reviews

    (2018)
  • H. Sun et al.

    Assessing the potential of random forest method for estimating solar radiation using air pollution index

    Energy Conversion and Management

    (2016)
  • S. Sun et al.

    A decomposition-clustering-ensemble learning approach for solar radiation forecasting

    Solar Energy

    (2018)
  • Chung J., C. Gulcehre, K. Cho, and Y. Bengio. (2014). 'Empirical evaluation of gated recurrent neural networks on...
  • S. Faizollahzadeh Ardabili et al.

    Computational intelligence approach for modeling hydrogen production: A review

    Engineering Applications of Computational Fluid Mechanics

    (2018)
  • C. Feng et al.

    Unsupervised clustering-based short-term solar forecasting

    IEEE Transactions on Sustainable Energy

    (2018)
  • R. Fermín et al.

    Predicting solar energy generation through artificial neural networks using weather forecasts for microgrid control

    Renewable Energy

    (2018)
  • F. Fotovatikhah et al.

    Survey of computational intelligence as basis to big flood management: Challenges, research directions and future work

    Engineering Applications of Computational Fluid Mechanics

    (2018)
  • L. Fu et al.

    A regional photovoltaic output prediction method based on hierarchical clustering and the mRMR criterion

    Energies

    (2019)
  • S. Ghimire et al.

    Deep solar radiation forecasting with convolutional neural network and long short-term memory network algorithms

    Applied Energy

    (2019)
  • X. Guo et al.

    Adaptive self-paced deep clustering with data augmentation

    IEEE Transactions on Knowledge and Data Engineering

    (2020)
  • Cited by (0)

    1

    ORCID: 0000-0002-4169-4438.

    2

    ORCID: 0000-0002-4271-6483.

    3

    ORCID: 0000-0003-0783-3585.

    4

    ORCID: 0000-0003-4786-7931.

    View full text