Ocean wave energy forecasting using optimised deep learning neural networks

Highlights

• We propose a new methodology for short-term wave forecasting.

• A Deep Neural Network structure is used to perform the forecasts.

• Moth-Flame Optimizer is used to “shape and fine-tune” sensitive hyperparameters.

• A broad spectrum of wave data from NDBC and ISDM operators were used for testing.

• Results revealed promising forecasting accuracy, especially in short-range horizons.

Abstract

Ocean renewable energy is a promising inexhaustible source of renewable energy, with an estimated harnessing potential of approximately 337 GW worldwide, which could re-shape the power generation mix. As with other sources of renewables, however, wave energy has an intermittent and irregular nature, which is a major concern for power system stability. Consequently, in order to integrate wave energy into power grids, it must be forecasted. This paper proposes using optimised deep learning neural networks to forecast the wave energy flux, and other wave parameters. In particular, we use moth-flame optimisation as the central decision-making unit to configure the deep neural network structure and the proper input data selection. Besides, the moth-flame optimisation algorithm was modified to improve its search space mechanisms. The forecasting skills are assessed using 13 datasets from locations across the Pacific and Atlantic coasts, and the Gulf of Mexico. The proposed optimised deep neural network performs well at all the sites, especially over short-term horizons, where it outperforms statistical and physics-based approaches.

Introduction

The focus of any green agenda is the consolidation of renewable power generation technologies as the primary source for electricity generation. In this regard, marine energy resources are a potentially large source of renewable energy, and estimates suggest that by 2050, up to 337 GW of installed capacity worldwide (of which 100 GW in Europe alone) could be available (Badcock-Broe et al., 2014). Besides, waves exhibit energy densities many times greater than do the solar and wind resources (Mendes et al., 2012). Another great example of this potential is the Australian coast, with its ocean renewable energy (ORE) resources far exceeding the current electricity demand (Hemer et al., 2018).

An essential stage to not only determine the economic viability of an ORE project but also in the technical side, i.e., when deciding upon the proper WEC mechanical structure and operational principles, is to correctly map and quantify the average wave energy available at specific locations. This is commonly achieved by analysing the buoy data and employing deep-water numerical models (using bathymetry, wind, and radar data as inputs). The set of measurable parameters typically includes significant wave height, mean and peak wave periods, and mean wave direction (Mendes et al., 2012; Uihlein and Magagna, 2016). The WEC power output is characterised over discretised ranges of these met-ocean parameters, particularly, the significant wave height and the energy period, forming what is known as the performance matrix (Hiles et al., 2016).

Moreover, wave conditions are known to exhibit variability over monthly, seasonal, inter-annual, and decadal time-scales (Cahill and Lewis, 2014). This variability is a consequence of various factors, such as fluctuations of the yearly wind index, climate indices, seasonality of wind speed, and solar irradiance. For example, there can be considerably more wind during winters than in summers, which is contrary to the case of solar energy.

The technological breakthroughs on wave farming pose special operational concerns from the perspective of utilities, as the ongoing power grid integration process will require compliance with a set of operational details. Another important issue is the question of predictability with regard to short-term trading on electricity markets, where generation imbalances are met with hefty penalties (Jeon and Taylor, 2016). In turn, the authors in (Reikard et al., 2015a) studied the costs of integrating the wave energy (over the Pacific Northwest zone) into the power grid, in conjunction with a wind and solar portfolio, outlining the reduced need for wave balancing reserves in relation to traditional wind and solar farms. This feature was also demonstrated when forecasting the ocean wave energy output in the coastal region of western Canada (Reikard et al., 2015b).

As previously observed in the case of wind and solar energy forecasts, the researchers have been looking for suitable solutions to accurately forecast the oceanographic parameters, with a special emphasis on wave power output and wave height. In the same fashion as for other time-series, the literature on this subject ranges from physics-based (numerical) models, classical statistical methods, soft-computing methods to hybrid methods. Of these, the physics-based (numerical) models typically exhibit a better accuracy over larger horizons (Reikard et al., 2015a) and perform well for longer intervals. The growth in the observations of sea-state parameters and the unceasing developments in modelling the dynamics of ocean waves and atmospheric interactions has been the key to massive improvements in the ability of physics-based models of different generations to estimate the sea-state parameters (Janssen, 2008). For instance, numerical forecasting experiments were carried out based on the (physical) wave model WW3 and were applied to the North Pacific and China Sea datasets (Bell and Kirtman, 2018, Zheng et al., 2016). Similarly, another third-generation model named WAM is also among the most popular choices.

With regard to statistical and soft-computing approaches (Hadadpour et al., 2014), achieved good results using a NN with specific input selection models. In turn (Reikard et al., 2011), considered 13 datasets to benchmark the wave model from the European Centre for Medium-range Weather Forecasts (ECMWF) against time-varying regression and a multi-layer perceptron. The latter performed better for short-term horizons, whereas the physics-based model performed better for lead times above 5 h with a narrower error range.

The same conclusions were drawn in (Reikard et al, 2015a, 2017) for a wide-range of locations (i.e., different locations/sea-state parameters). Furthermore, the recurrent NNs are also amongst the fittest options for time-series prediction. Using this approach, authors in (Desouky and Abdelkhalik, 2019; Sadeghifar et al., 2017) employed nonlinear autoregressive exogenous (NARX) networks for wave prediction in the South Caspian and two locations near the Hawaiian coast, respectively.

To enhance the performance of the predictions, a common practice is to apply wavelet transformation or wavelet neural networks to expose different characteristics of the time-series (Özger, 2010; Prahlada and Deka, 2015).

Furthermore, recently trending approaches have continued to cover soft-computing methods, such as support vector machines, extreme learning machines, sequential learning NN, genetic fuzzy systems, and machine learning applications (Berbić et al., 2017; Cornejo-Bueno et al., 2018; Fernández et al., 2015; James et al., 2018; Kumar et al, 2017, 2018). In (Akbarifard and Radmanesh, 2018) the Symbiotic Organisms Search (SOS) algorithm was used to fine-tune the coefficients of statistical forecasting models in order to predict the wave height in the Caspian Sea. Moreover, in (Duran-Rosal et al., 2017) an artificial neural network classifier with hybrid basis functions is trained with a multi-objective evolutionary algorithm (MOEA) in order to predict Extreme Significant Wave Height segment.

Following the success of these approaches, a new methodology based on a Deep NN as a preferred forecasting engine, is presented in this work. This methodology's sensitive aspects with regard to architecture and input data selection are automatically fine-tuned by an improved Moth-Flame Optimisation (MFO) algorithm. This methodology is validated for a variety of periods and horizons using an extensive set of samples measured from 13 locations along the US and Canadian coasts.

The remainder of this paper is organised in the following manner: Section II provides a background review of major topics covered in this work, namely the theory behind the sea state parameters, a brief review of artificial NNs and the improved version of the MFO algorithm. Section III follows by presenting meaningful characteristics of the chosen 13 datasets (locations). A detailed description of the proposed methodology is presented in Section IV. Section V presents and discusses the test results of the proposed methodology. Lastly, Section VI presents the key conclusions of this work.