Elsevier

Renewable Energy

Volume 105, May 2017, Pages 569-582
Renewable Energy

Review
Machine learning methods for solar radiation forecasting: A review

https://doi.org/10.1016/j.renene.2016.12.095Get rights and content

Highlights

  • Overview of forecasting methods of solar irradiation using machine learning approaches.

  • Performance ranking of such methods is complicated.

  • ANN and ARIMA methods are equivalent in term of quality of prediction.

  • Predictor ensemble methodology is always better than simple predictors.

  • SVM, regression trees and random forests, as the results given are very promising.

Abstract

Forecasting the output power of solar systems is required for the good operation of the power grid or for the optimal management of the energy fluxes occurring into the solar system. Before forecasting the solar systems output, it is essential to focus the prediction on the solar irradiance. The global solar radiation forecasting can be performed by several methods; the two big categories are the cloud imagery combined with physical models, and the machine learning models. In this context, the objective of this paper is to give an overview of forecasting methods of solar irradiation using machine learning approaches. Although, a lot of papers describes methodologies like neural networks or support vector regression, it will be shown that other methods (regression tree, random forest, gradient boosting and many others) begin to be used in this context of prediction. The performance ranking of such methods is complicated due to the diversity of the data set, time step, forecasting horizon, set up and performance indicators. Overall, the error of prediction is quite equivalent. To improve the prediction performance some authors proposed the use of hybrid models or to use an ensemble forecast approach.

Introduction

An electrical operator should ensure a precise balance between the electricity production and consumption at any moment. This is often very difficult to maintain with conventional and controllable energy production system, mainly in small or not interconnected (isolated) electrical grid (as found in islands). Many countries nowadays consider using renewable energy sources into their electricity grid. This creates even more problems as the resource (solar radiation, wind, etc.) is not steady. It is therefore very important to be able to predict the solar radiation effectively especially in case of high energy integration [1].

One of the most important challenge for the near future global energy supply will be the large integration of renewable energy sources (particularly non-predictable ones as wind and solar) into existing or future energy supply structure. An electrical operator should ensure a precise balance between the electricity production and consumption at any moment. As a matter of fact, the operator has often some difficulties to maintain this balance with conventional and controllable energy production system, mainly in small or not interconnected (isolated) electrical grid (as found in islands). The reliability of the electrical system then become dependent on the ability of the system to accommodate expected and unexpected changes (in production and consumption) and disturbances, while maintaining quality and continuity of service to the customers. Then, the energy supplier must manage the system with various temporal horizons (see Fig. 1).

The integration of renewable energy into an electrical network intensifies the complexity of the grid management and the continuity of the production/consumption balance due to their intermittent and unpredictable nature [1], [2]. The intermittence and the non-controllable characteristics of the solar production bring a number of other problems such as voltage fluctuations, local power quality and stability issues [3], [4]. Thus forecasting the output power of solar systems is required for the effective operation of the power grid or for the optimal management of the energy fluxes occurring into the solar system [5]. It is also necessary for estimating the reserves, for scheduling the power system, for congestion management, for the optimal management of the storage with the stochastic production and for trading the produced power in the electricity market and finally to achieve a reduction of the costs of electricity production [1], [3], [6], [7]. Due to the substantial increase of solar power generation the prediction of solar yields becomes more and more important [8]. In order to avoid large variations in renewable electricity production it is necessary to include also the complete prediction of system operation with storage solutions. Various storage systems are being developed and they are a viable solution for absorbing the excess power and energy produced by such systems (and releasing it in peak consumption periods), for bringing very short fluctuations and for maintaining the continuity of the power quality. These storage options are usually classified into three categories:

  • -

    Bulk energy storage or energy management storage media is used to decouple the timing of generation and consumption.

  • -

    Distributed generation or bridging power - this method is used for peaks shaving - the storage is used for a few minutes to a few hours to assure the continuity of service during the energy sources modification.

  • -

    The power quality storage with a time scale of about several seconds is used only to assure the continuity of the end use power quality.

Table 1 shows these three categories and their technical specifications. As shown, every type of storage is used in different cases to solve different problems, with different time horizon and quantities of energy.

Table 1 shows that the electricity storage can be widely used in a lot of cases and applications as a function of the time of use and the power needs of the final user. Finally, it shows that the energy storage acts at various time levels and their appropriate management requires the knowledge of the power or energy produced by the solar system at various horizons: very short or short for power quality category to hourly or daily for bulk energy storages. Similarly, the electrical operator needs to know the future production (Fig. 1) at various time horizons from one to three days, for preparing the production system and to some hours or minutes for planning the start-up of power plants (Table 2). Starting a power plant needs between 5 min for a hydraulic one to 40 h for a nuclear one. Moreover, the rise in power of the electrical plants is sometimes low, thus for an effective balance between production and consumption an increase of the power or a starting of a new production needs to be anticipated sometimes well in advance.

Furthermore, the relevant horizons of forecast can and must range from 5 min to several days as it was confirmed by Diagne et al. [6]. Elliston and MacGill [10] outlined the reasons to predict solar radiation for various solar systems (PV, thermal, concentrating solar thermal plant, etc.) insisting on the forecasting horizon. It therefore seems apparent that the time-step of the predicted data may vary depending on the objectives and the forecasting horizon. All these reasons show the importance of forecasting, whether in production or in consumption of energy. The need for forecasting lead to the necessity to use effective forecasting models. In the next section the various available forecasting methodologies are presented.

The solar power forecasting can be performed by several methods; the two big categories are the cloud imagery combined with physical models, and the machine learning models. The choice for the method to be used depends mainly on the prediction horizon; actually all the models have not the same accuracy in terms of the horizon used. Various approaches exist to forecast solar irradiance depending on the target forecasting time. The literature classifies these methods in two classes of techniques:

  • -

    Extrapolation and statistical processes using satellite images or measurements on the ground level and sky images are generally suitable for short-term forecasts up to 6 h. This class can be divided in two sub-classes, in the very short time domain called “Now-casting’’ (0–3 h), the forecast has to be based on extrapolations of real-time measurements [5]; in the Short-Term Forecasting (3–6 h), Numerical Weather Prediction (NWP) models are coupled with post-processing modules in combination with real-time measurements or satellite data [5], [11].

  • -

    NWP models able to forecast up to two days ahead or beyond [12], [13] (up to 6 days ahead [13]). These NWP models are sometimes combined with post-processing modules and satellite information are often used [2].

Fig. 2a and b [6], [14] summarize the existing methods versus the forecasting horizon, the objective and the time step.

The NWP models predict the probability of local cloud formation and then predict indirectly the transmitted radiation using a dynamic atmosphere model. The extrapolation or statistical models analyze historical time series of global irradiation, from satellite remote sensing [15] or ground measurements [16] by estimating the motion of clouds and project their impact in the future [6], [13], [17]. Hybrid methods can improve some aspects of all of these methods [6], [14]. The statistical approach allows to forecast hourly solar irradiation (or at a lower time step) and NWP models use explanatory variables (mainly cloud motion and direction derived from atmosphere) to predict global irradiation N-steps ahead [15]. Very good overviews of the forecasting methods, with their limitations and accuracy can be found in Refs. [1], [5], [6], [10], [12], [14], [18]. Benchmarking studies were performed to assess the accuracy of irradiance forecasts and compare different approaches of forecasting [8], [13], [17], [19], [20], [21]. Moreover, the accuracy evaluation parameters are often different; some parameters such as correlation coefficient and root mean square error are often used, but not always adapted to compare the model performance. Thus the time period used for evaluating the accuracy varies widely. Some of them analyzed the model accuracy over a period of one or several years, whereas some others over a period of some weeks introducing a potential seasonal bias. In these conditions, it is not easy to make comparisons and the accuracy of the results produced, as shown in this paper, must be carefully evaluated in selecting the right method to use. As part of COST Action ES1002. (European Cooperation in Science and Technology) [22] on Weather Intelligence for Renewable Energies (WIRE) a literature review on the forecasting accuracy applied to renewable energy systems mainly solar and wind is carried out. In this paper an overview on the various methodologies available for solar radiation prediction based on machine learning is presented. A lot of review papers are available, but it is very rare to find a paper which is totally dedicated to the machine learning methods and that some recent prediction models like random forest, boosting or regression tree be integrated. In the next section the different methodologies used in the literature to predict global radiation and the parameters used for estimating the model performances are presented.

Section snippets

Machine learning methods

Machine learning is a subfield of computer science and it is classified as an artificial intelligence method. It can be used in several domains and the advantage of this method is that a model can solve problems which are impossible to be represented by explicit algorithms. In Ref. [23] the reader can find a detailed review of some machine learning and deterministic methods for solar forecasting. The machine learning models find relations between inputs and outputs even if the representation is

Evaluation of model accuracy

Evaluation, generally, measures how good something is. This evaluation is used at various steps of the model development as for example during the evaluation of the forecasting model itself (during the training of a statistical model for example), for judging the improvement of the model after some modifications and for comparing various models. As previously mentioned, this performance comparison is not easy for various reasons such as different forecasted time horizons, various time scale of

Machine learning forecasters’ comparison

Before presenting the results related to the machine learning method in order to predict the global radiation, Fig. 6 shows the number of times the term ANN, machine learning and SVM/SVR are referenced in the five main journals of solar energy prediction (Solar Energy, Energy, Applied Energy, Renewable Energy and Energy Conversion and Management).

All three terms are more and more used in literature. It can be seen the ANN is the mostly used method in global radiation forecasting.

Conclusions and outlook

As shown in the present paper, many methods and types of methods are available. There are a lot of methods to estimate the solar radiation, some are often used (ANN, ARIMA, naive methods), others begin to be used (SVM, SVR, k-mean) more frequently and other are rarely used (boosting, regression tree, random forest, etc.). In some cases, one is the best and in others it is the reverse. As a conclusion, it can be said that the ANN and ARIMA methods are equivalent in term of quality of prediction

Acknowledgement

This work was carried out in the framework of the Horizon 2020 project (H2020-LCE-2014-3 - LCE-08/2014 - 646529) TILOS “Technology Innovation for the Local Scale, Optimum Integration of Battery Energy Storage”. The authors would like to acknowledge the financial support of the European Union.

References (102)

  • Z. Şen et al.

    Solar irradiation estimation by monthly principal component analysis

    Energy Convers. Manag.

    (2008)
  • M. Zarzo et al.

    Modeling the variability of solar radiation data among weather stations by means of principal components analysis

    Appl. Energy

    (2011)
  • C. Paoli et al.

    Forecasting of preprocessed daily solar radiation time series using neural networks

    Sol. Energy

    (2010)
  • J.G. De Gooijer et al.

    25 years of time series forecasting

    Int. J. Forecast

    (2006)
  • H.-Y. Cheng

    Hybrid solar irradiance now-casting by fusing Kalman filter and regressor

    Renew. Energy

    (2016)
  • S. Kalogirou

    Artificial neural networks in renewable energy systems applications: a review

    Renew. Sustain. Energy Rev.

    (2001)
  • C. Voyant et al.

    Multi-horizon solar radiation forecasting for Mediterranean locations using time series models

    Renew. Sustain. Energy Rev.

    (2013)
  • M. Chaabene et al.

    Neuro-fuzzy dynamic model with Kalman filter to forecast irradiance and temperature for solar energy systems

    Renew. Energy

    (2008)
  • A. Troncoso et al.

    Local models-based regression trees for very short-term wind speed prediction

    Renew. Energy

    (2015)
  • H.T.C. Pedro et al.

    Nearest-neighbor methodology for prediction of intra-hour global horizontal and direct normal irradiances

    Renew. Energy

    (2015)
  • G. Zhang

    Time series forecasting using a hybrid ARIMA and neural network model

    Neurocomputing

    (2003)
  • C. Bouveyron et al.

    Model-based clustering of high-dimensional data: a review

    Comput. Stat. Data Anal.

    (2014)
  • L. Gaillard et al.

    Data-driven performance evaluation of ventilated photovoltaic double-skin facades in the built environment

    Energy Procedia

    (2015)
  • Y. Gala et al.

    Hybrid machine learning forecasting of solar radiation values

    Neurocomputing

    (2016)
  • M.P. Almeida et al.

    PV power forecast using a nonparametric PV model

    Sol. Energy

    (2015)
  • J. Huang et al.

    An analytical comparison of four approaches to modelling the daily variability of solar irradiance using meteorological records

    Renew. Energy

    (2014)
  • J.R. Trapero et al.

    Short-term solar irradiation forecasting based on dynamic harmonic regression

    Energy

    (2015)
  • P. Bacher et al.

    Online short-term solar power forecasting

    Sol. Energy

    (2009)
  • M. David et al.

    Probabilistic forecasting of the solar irradiance with recursive ARMA and GARCH models

    Sol. Energy

    (2016)
  • A. Mellit et al.

    Artificial intelligence techniques for sizing photovoltaic systems: a review

    Renew. Sustain. Energy Rev.

    (2009)
  • Y. Kemmoku et al.

    Daily insolation forecasting using a multi-stage neural network

    Sol. Energy

    (1999)
  • G.P. Podestá et al.

    Estimating daily solar radiation in the argentine pampas

    Agric. For. Meteorol

    (2004)
  • G.K.F. Tso et al.

    Predicting electricity energy consumption: a comparison of regression analysis, decision tree and neural networks

    Energy

    (2007)
  • R. Marquez et al.

    Forecasting of global and direct solar irradiance using stochastic learning methods, ground experiments and the NWS database

    Sol. Energy

    (2011)
  • A. Moreno et al.

    Mapping daily global solar irradiation over Spain: a comparative study of selected approaches

    Sol. Energy

    (2011)
  • S. Ben Taieb et al.

    A review and comparison of strategies for multi-step ahead time series forecasting based on the NN5 forecasting competition

    Expert Syst. Appl.

    (2012)
  • H. Long et al.

    Analysis of daily solar power prediction with data-driven approaches

    Appl. Energy

    (2014)
  • S. Salcedo-Sanz et al.

    Daily global solar radiation prediction based on a hybrid Coral Reefs Optimization – extreme Learning Machine approach

    Sol. Energy

    (2014)
  • M. Zamo et al.

    A benchmark of statistical regression methods for short-term forecasting of photovoltaic electricity production, part I: deterministic forecast of hourly production

    Sol. Energy

    (2014)
  • M. Lazzaroni et al.

    Models for solar radiation prediction based on different measurement sites

    Measurement

    (2015)
  • J. Cao et al.

    Study of hourly and daily solar irradiation forecast using diagonal recurrent wavelet neural networks

    Energy Convers. Manag.

    (2008)
  • M. Bouzerdoum et al.

    A hybrid model (SARIMA–SVM) for short-term power forecasting of a small-scale grid-connected photovoltaic plant

    Sol. Energy

    (2013)
  • Y. Chu et al.

    Hybrid intra-hour DNI forecasts with sky image processing enhanced by stochastic learning

    Sol. Energy

    (2013)
  • S.K. Aggarwal et al.

    Solar energy prediction using linear and non-linear regularization models: a study on AMS (American Meteorological Society) 2013–14 solar energy prediction contest

    Energy

    (2014)
  • I. Bilionis et al.

    Data-driven model for solar irradiation based on satellite observations

    Sol. Energy

    (2014)
  • J. Wu et al.

    Prediction of solar radiation with genetic approach combing multi-model framework

    Renew. Energy

    (2014)
  • Y. Chu et al.

    Real-time forecasting of solar irradiance ramps with smart image processing

    Sol. Energy

    (2015)
  • M. De Felice et al.

    Short-term predictability of photovoltaic production over Italy

    Renew. Energy

    (2015)
  • B. Espinar et al.

    Photovoltaic Forecasting: a State of the Art

    (2010)
  • A. Moreno-Munoz et al.

    Very short term forecasting of solar radiation

  • Cited by (1225)

    View all citing articles on Scopus
    View full text