Data-driven symbolic ensemble models for wind speed forecasting through evolutionary algorithms
Introduction
The worst energy crisis in Brazil’s history, which occurred in 2001, added to the drought that has been affecting the northeast region for the last seven years, have been leading to substantial changes in the Brazilian energy matrix, which currently includes different types of renewable energy sources other than hydropower. In northeastern Brazil, the main source of energy is the wind power available in the region, which currently provides more than 50% of the daily energy production. The northeast of Brazil is among the best regions in the world for wind farming, with strong, persistent, and well-behaved winds [1], [2], [3].
The cubic relationship between the wind speed and theamount of energy produced by a wind turbine makes wind speed forecasting an important issue for decision-making in operations management of the wind power system in order to avoid failures on the power grid. In this way, one of the major research challenges in the context of wind power generation systems is to provide more accurate and reliable wind speed and power forecasts. There has not been much research regarding wind direction forecasts because in practice the turbines are able to automatically align their blades to the prevailing wind direction.
A large literature review on the topic of wind speed and power forecasting is given by Giebel et al. [4], who discuss a collection of published research relevant to the wind power management system. The current statistical methods adopted include autoregressive models, moving average models, autoregressive moving average models, autoregressive integrated moving average models, and Kalman filters [5], [6], [7]. From the machine learning area, artificial neural networks [8], [9], [10] are the most widely used approach, although fuzzy logic, support vector machines, neuro-fuzzy networks [11], and hybrid models [5], [12], [13] have also been tested. Ensemble weather forecast techniques for post-processing wind speed forecasts consist in combining multiple numerical weather prediction (NWP) models using fixed linear approaches based on simple average or performance-based weighted average, such as the MASTER super model ensemble system (MSMES) [14] and the Bayesian model averaging [15], [16].
However, the aforementioned techniques have some well-known limitations, such as linear representations with fixed structures, black-box frameworks, and the improvement of only a single solution during the optimization process. Despite these limitations, little has been experienced beyond that. For instance, there have been very few studies concerning nature-inspired algorithms in the wind-forecasting literature. Until recently, nature-inspired algorithms, such as artificial bee colony algorithm [17], particle swarm optimization [18], [19], [20], and differential evolution [20] have been used exclusively for parameter optimization rather than for solving the wind-forecasting problem directly. To the best of our knowledge, the powerful evolutionary algorithm known as grammatical evolution has never been explored as an alternative tool to tackle the challenges related to wind power generation systems. An example of successful application of grammatical evolution in ensemble forecast problems may be found in Dufek et al. [21], where the authors focused on ensemble forecasts for daily rainfall amount at 317 locations in Brazil.
In contrast to single-solution algorithms which tend to get stuck in local optima, grammatical evolution (GE) is a population-based, stochastic, global optimization algorithm that, together with its high degree of parallelism, allows for a better exploration of the whole search space, which in turn increases the probability of finding the global optimum. The white-box nature of the GE models provides many advantages and further applications over black-box approaches, such as: (i) the direct extraction of knowledge from the resulting symbolic solutions and (ii) the validation of the GE solutions’ structures by experts. In theory, the GE-based modeling can be defined as a general-purpose technique, as opposed to specific-purpose techniques, such as the conventional statistical methods and the spatial correlation models. In other words, the search space of the GE-based modeling is an extension to those of the specific-purpose techniques as it enables the use of not only standard arithmetic operators and NWP models, but also several other linear and non-linear operators (e.g. mathematical, logical, relational and conditional) as well as input attributes according to the problem at hand. The GE-based modeling can evolve expressions of arbitrary structure guided by formal grammars. In the context of wind speed forecasting, it is capable of, for example, (i) correcting an NWP model; (ii) optimizing a multi-model ensemble forecast; or (iii) potentially recovering the mathematical expressions associated with spatial–temporal models. An introduction to GE is given in Section 2.1. For more details the reader is referred to [22], [23].
The main contribution of this paper to the strategic planning of the wind power sector consists of (i) the proposal of a GE-based data-driven modeling in order to increase the accuracy and confidence of near-surface hourly wind speed point forecasts; (ii) a case study for one- to three-day-ahead forecasts at five locations in northeastern Brazil; (iii) an investigation into the influence of feature selection and execution time on the GE forecast accuracy; (iv) providing information about the spatial and temporal variability of the GE forecast accuracy; (v) comparing the accuracy of the GE solutions with those obtained from five other approaches; and (vi) presenting a way of extracting knowledge and insights from the human-interpretable (“white-box”) solutions given by GE through sensitivity analysis. The methodology designed to deal with the wind-forecasting problem is based on three steps. The first step includes the feature selection from a pool of possible predictors, containing individual and ensemble numerical weather forecasts, and historical data from the target location and its vicinity. Next, it is the execution of the GE algorithm making use of multi- and many-core parallelism. In the third step, the best solutions are chosen and analyzed.
Section snippets
Grammatical evolution
Grammatical evolution (GE) is an algorithm that automatically builds functional structures (“programs”) by means of an iterative optimization process inspired by the evolutionary principle of natural selection [22]. It is essentially a genetic algorithm [24], sharing the same representation and breeding process carried out over a number of generations, but equipped with a more sophisticated mechanism to map the genotype space (population of individuals encoded as bit-arrays) into the phenotype
Results and discussion
In this section several GE experiments are analyzed from two perspectives: technique- and problem-centered analyses.
The technique-centered analysis studies the influence of feature selection (Section 3.1) and execution time (Section 3.2) on the accuracy of the 24-, 48- and 72-h forecasts of near-surface hourly wind speed given by GE over northeastern Brazil. Thereafter, in Section 3.3 the GE-based data-driven modeling is compared in terms of forecast error with MSMES and four other approaches:
Conclusions and future work
This paper presented a GE-based data-driven modeling of the 24-, 48- and 72-h forecasts of near-surface hourly wind speed at five locations over northeastern Brazil. Several GE experiments were conducted under two different points of view: (i) a technique-centered analysis which concerns the efficiency, effectiveness, scalability, and robustness of the GE-based data-driven modeling; and (ii) a problem-centered analysis which focuses on understanding the regression problem of predicting wind
Declaration of Competing Interest
No author associated with this paper has disclosed any potential or pertinent conflicts which may be perceived to have impending conflict with this work. For full disclosure statements refer to https://doi.org/10.1016/j.asoc.2019.105976.
Acknowledgments
The authors would like to thank the support provided by CNPq, Brazil (grants 312337/2017-5, 502836/2014-8 and 300458/2017-7), FAPEMIG, Brazil (grant APQ-03414-15), EU H2020 Programme and MCTI/RNP–Brazil under the HPC4E Project (grant 689772).
References (36)
- et al.
Wind energy assessment and wind farm simulation in Triunfo – Pernambuco, Brazil
Renew. Energy
(2010) - et al.
Comparison of two new ARIMA-ANN and ARIMA-Kalman hybrid methods for wind speed prediction
Appl. Energy
(2012) - et al.
ARMA based approaches for forecasting the tuple of wind speed and direction
Appl. Energy
(2011) - et al.
Wind power prediction using deep neural network based meta regression and transfer learning
Appl. Soft Comput.
(2017) - et al.
Analysis of wind power generation and prediction using ANN: A case study
Renew. Energy
(2008) Artificial neural networks in renewable energy systems applications: A review
Renew. Sustain. Energy Rev.
(2001)- et al.
A generalized dynamic fuzzy neural network based on singular spectrum analysis optimized by brain storm optimization for short-term wind speed forecasting
Appl. Soft Comput.
(2017) - et al.
An innovative hybrid approach for multi-step ahead wind speed prediction
Appl. Soft Comput.
(2019) - et al.
Short-term wind speed forecasting based on a hybrid model
Appl. Soft Comput.
(2013) - et al.
Short-term wind power forecasting using adaptive neuro-fuzzy inference system combined with evolutionary particle swarm optimization, wavelet transform and mutual information
Renew. Energy
(2015)
Particle swarm optimization for construction of neural network-based prediction intervals
Neurocomputing
Short-term wind power forecasting using evolutionary algorithms for the automated specification of artificial intelligence models
Int. J. Forecast.
Application of evolutionary computation on ensemble forecast of quantitative precipitation
Comput. Geosci.
Assessment of wind resources in two parts of northeast Brazil with the use of numerical models
Meteorol. Appl.
Wind resource evaluation in São João do Cariri (SJC) – Paraíba, Brazil
Renew. Sustain. Energy Rev.
The State-Of-The-Art in Short-Term Prediction of Wind Power: A Literature Overview
Modelling and forecasting wind speed intensity for weather risk management
Comput. Statist. Data Anal.
The Master Super Model Ensemble system (MSMES)
Cited by (7)
Simultaneous prediction for multiple source–loads based sliding time window and convolutional neural network
2022, Energy ReportsCitation Excerpt :When carrying out prediction, uncertainty, temporal properties and relevance, prediction accuracy are extensively concerned issues. For improvement of prediction accuracy and considering wind speed randomness and uncertainty, Dufek et al. (2020) proposed a method based on vibrational mode decomposition, drosophila optimization algorithm, and autoregressive integrated moving average model. Based on a Weight-Varying combination forecast mode, Yu et al. (2019) developed a novel two-stage model to quantify the uncertainty of PV power by the prediction interval.
Point and interval forecasting of ultra-short-term wind power based on a data-driven method and hybrid deep learning model
2022, EnergyCitation Excerpt :In summary, there still exist research gaps in the current academic community and engineering practice on multi-dimensional data cleaning and feature mining technology of wind farm data space; intelligent decomposition and noise reduction techniques of wind power sequence; and the combination of advanced DL forecasting and interval forecasting models. To improve the forecasting accuracy, WPF based on data-driven ideas and advanced DL algorithms has become a research hotspot [24,33,58]. Extracting hidden features from massive high-dimensional and multi-modal data space is not only the advantage of DL models, but also the new development trend of AI.
Multi-step wind speed forecasting model based on wavelet matching analysis and hybrid optimization framework
2020, Sustainable Energy Technologies and AssessmentsCitation Excerpt :Many models have been proposed to achieve high precise wind speed forecasting. These models can be divided into the physical models [5] and the statistical models [6]. The physical models simulate the wind base on the physical laws and usually take terrain and meteorological factors into consideration [7].
Development of NCL equivalent serial and parallel python routines for meteorological data analysis
2022, International Journal of High Performance Computing ApplicationsA holistic review on energy forecasting using big data and deep learning models
2021, International Journal of Energy Research