Ocean wave energy forecasting using optimised deep learning neural networks
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.
Section snippets
Sea state parameters and wave power density
As mentioned previously, the abundant energy densities (energy transport) associated with ocean waves, tied to the fact that oceans cover more than 70% of the earth's surface, constitute a massive renewable energy source. Recurrently, the term ‘wave’ is misinterpreted, which in its essence characterises how energy is transferred, in this case through water as medium. The ocean waves are a result of wind interaction with the water surface (energy transfer) (Khan et al., 2017). This ultimately
Wave data
In the same fashion as in (Reikard et al., 2011), buoy measurements and derived data consist of meteorological hourly datasets from 11 buoys from the US National Data Buoy Centre (NDBC) and two (South Brooks and East Dellwood) from the Canadian Integrated Science Data Management (ISDM). These 13 locations enable a broader validation, which implies that conclusions are not constrained to a particular geographic zone or a particular sea-state condition.
These datasets include quality controlled
Proposed methodology
As we saw earlier, the use of soft-computing approaches is nowadays an appealing solution to a wide variety of domains; particularly deep learning architectures provide a great framework to solve hard classification and prediction problems. Regarding this last set of problems, deep learning architectures despite their ‘black-box’ nature are effective forecasting engines, exhibiting impressive self-learning capabilities, when trained with a comprehensive (representative) dataset. Deep learning
Results
The proposed methodology is tested using the data presented in Table 1 (13 datasets). In other words, variables , , and are forecast for each location with short range lead times of 1, 2, 3, 6, and 12 h. Moreover, the wave energy flux is not only forecasted directly (i.e., the DNN output provides the forecast), but also indirectly by using the and forecasts, using Eq. (1), under a set of assumptions (that fit a broad range of conditions).
For validation purpose, four months of
Conclusions
With a need to respond to the existential threat posed by climate change, societies in general, and electric power systems, in particular, are called upon to make a contribution to environmental protection, commensurate with their responsibilities. To this end, a tremendous effort was put into the increase of the share of renewables in the conventional power generation portfolio. In addition, the maturation of technologies such as photovoltaic (PV) cells, solar, and wind, brings in additional
CRediT authorship contribution statement
P.M.R. Bento: Conceptualization, Methodology, Software, Validation, Investigation, Writing - original draft, Writing - review & editing. J.A.N. Pombo: Conceptualization, Methodology, Software, Validation, Investigation. R.P.G. Mendes: Conceptualization, Methodology, Investigation. M.R.A. Calado: Conceptualization, Methodology, Validation, Investigation, Writing - review & editing. S.J.P.S. Mariano: Conceptualization, Methodology, Software, Validation, Investigation, Supervision, Resources,
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
The authors would like to express their gratitude to Doctor Jean-Raymond Bidlot for the provided data and valuable insights on the subject, that made this work possible. This gratitude is extended to the Industry Statistician, Gordon Reikard, for the initial guidance and kind availability.
P.M.R. Bento gives his special thanks to the Fundação para a Ciência e a Tecnologia (FCT), Portugal, for the Ph.D. Grant (SFRH/BD/140371/2018).
References (46)
- et al.
Approximation of the joint probability density of wave steepness and height with a bivariate gamma distribution
Ocean Eng.
(2016) - et al.
Application of neural networks and support vector machine for significant wave height prediction
Oceanologia
(2017) - et al.
Significant wave height and energy flux estimation with a Genetic Fuzzy System for regression
Ocean Eng.
(2018) - et al.
Wave prediction using wave rider position measurements and NARX network in wave energy conversion
Appl. Ocean Res.
(2019) - et al.
A multi-layer perceptron approach for accelerated wave forecasting in Lake Michigan
Ocean Eng.
(2020) - et al.
Significant wave height and energy flux range forecast with machine learning classifiers
Eng. Appl. Artif. Intell.
(2015) - et al.
Perspectives on a way forward for ocean renewable energy in Australia
Renew. Energy
(2018) - et al.
A machine learning framework to forecast wave conditions
Coast. Eng.
(2018) Progress in ocean wave forecasting
J. Comput. Phys.
(2008)- et al.
Short-term density forecasting of wave energy using ARMA-GARCH models and kernel density estimation
Int. J. Forecast.
(2016)
Wave height forecasting in dayyer, the Persian Gulf
Ocean Eng.
Review of ocean tidal, wave and thermal energy technologies
Renew. Sustain. Energy Rev.
Ocean wave height prediction using ensemble of Extreme Learning Machine
Neurocomputing
Regional ocean wave height prediction using sequential learning neural networks
Ocean Eng.
Wave energy potential in Portugal–Assessment based on probabilistic description of ocean waves parameters
Renew. Energy
Moth-flame optimization algorithm: a novel nature-inspired heuristic paradigm
Knowl. Base Syst.
Significant wave height forecasting using wavelet fuzzy logic approach
Ocean Eng.
Forecasting, hindcasting and feature selection of ocean waves via recurrent and sequence-to-sequence networks
Ocean Eng.
Forecasting of time series significant wave height using wavelet decomposed neural network
Aquat. Procedia
Daily long-term traffic flow forecasting based on a deep neural network
Expert Syst. Appl.
Forecasting ocean wave energy: the ECMWF wave model and time series methods
Ocean Eng.
Wave energy worldwide: simulating wave farms, forecasting, and calculating reserves
Int. J. Mar. Energy
Combining wave energy with wind and solar: short-term forecasting
Renew. Energy
Cited by (64)
A deep hybrid network for significant wave height estimation
2024, Ocean ModellingA short-term wave energy forecasting model using two-layer decomposition and LSTM-attention
2024, Ocean Engineering