The ongoing large scale adoption of wind power increases the associated risks related to the variability. An essential way to mitigate these risks for a utility company is to forecast production accurately. This study aims to create insight into the potential of deep learning models for both forecast quality and value on the ultra-short-term wind power forecasting (UST-WPF) horizon. The status quo at Eneco, a Dutch utility company, for UST-WPF is a numerical weather prediction (NWP) based model with a rudimentary ultra-short-term (UST) correction with real-time power data. The methodology followed during this thesis was the development of four UST-WPF models for Princess Amalia Wind Farm (PAWP) with a 16 programme time unit (PTU) forecast horizon and a forecast frequency of 1 PTU. Both model 1 and model 2 only use real-time data and are based on a multilayer perceptron (MLP) and a long short-term memory (LSTM) architecture, respectively. The other two proposed models are a multivariate combination of these two respective models with the Eneco model. The accuracy of the four proposed models was compared to two benchmark models: a Persistence and the Eneco model. Additionally, a novel framework was designed to evaluate the forecast value relative to the Eneco model on a variable forecast horizon. This study indicates that the proposed deep learning models can contribute both in quality and value up to 9 PTUs ahead.

The use of deep learning in global weather forecasting has shown significant promise in improving both forecasting accuracy and speed. Traditional numerical weather prediction models have gradually improved forecasting skills but at the cost of increased computational complexity. In contrast, new deep learning models, trained directly on reanalysis data, have demonstrated significant gains in forecasting accuracy, achieving competitive levels of performance. However, the potential of deep learning in predicting spatiotemporal chaotic systems, such as weather patterns, remains unexplored. To address this gap, we investigate the efficacy of a data-driven Fourier neural operator Markovian forecaster to replicate the intrinsic predictability and the characteristic Lyapunov spectrum of the Kuramoto-Sivashinsky system. FNO reproduces intrinsic predictability and precisely estimates the characteristic Lyapunov spectrum, even with a small dataset. They cannot represent one of the invariant symmetries, a zero characteristic Lyapunov exponent in the spectrum. Our findings suggest that deep learning can not only enhance the speed and accuracy of traditional numerical forecasting models but also replicate the weather's chaotic nature. This has significant implications for generating large ensembles and improving overall probabilistic forecasts. The research is limited to a deterministic system defined on a single process time scale surrogated by FNO. The future merits a similar study to investigate if FNO models can perform similarly on larger, stochastic, coupled, and/or multiple time-scale spatiotemporal chaotic systems.

With the growing global drive to act up on climate change, the adoption of renewable energy sources such as solar photovoltaic (PV) is continuously increasing. This crucial shift poses many economic and environmental benefits, however the variability in solar PV generation may also threaten the stability of our power grid and energy supply. The reliable prediction of this fluctuating power resource on various time scales has been identified as a crucial technology for the continuous massive adoption of solar PV. This study concentrates on the application of convolutional neural networks (CNN) and Long Short Term Memory (LSTM) to process real-time data sources in spatially aggregated solar PV power forecast for Germany, with specifically a forecast horizon of 3 hours and 15-minute interval. Two models are designed to be applicable in a real-time operational setting with a short forecast lag: (1) A LSTM network that leverages on the latest solar PV generation data and a NWP based day-ahead power forecast, and (2) a CNN-LSTM network designed to utilize the latest satellite images and a NWP based day-ahead power forecast. The accuracy of the forecast models are evaluated using one year of solar PV power generation data in Germany (January 2020 through December 2020), and are compared to a persistence model and a NWP based day-ahead and intra-day power forecast provided by the German transmission system operators. The empirical results show that the two proposed models perform equal or better than the benchmark models. An implication for power trading practices is that deep learning models, such as LSTM and CNN-LSTM, shows to be a promising forecasting technique which deserves a place in a comprehensive solar PV power forecasting toolbox.

The increasing penetration of weather-dependent energy sources brings additional challenges to the operation of the power system. Wind power forecasting is a valuable resource for these power operators: a tool that aids the decision-making process and facilitates risk management. On the other hand, the progress of machine learning and their success in different fields attracted research into wind power probabilistic forecasting. The objective of the research was to enhance wind power forecasting through data-driven or machine learning models, proposing a new framework as an alternative to the existing ones, validated in different climate regions using grid-like topology data (weather images). The project focused on the European Energy Markets 2020 Conference (EEM20) Forecasting Competition. The setting emulates the day-ahead electricity market and the location consisted of all four electricity price regions in Sweden, where three main sets of data were provided. Three main challenges are present in this competition, namely the large volume of data (the curse of dimensionality), the lack of information regarding wind turbine availability, and the growing evolution of wind installed capacity between data sets. The methodology to create a framework consisted of evaluating different machine learning approaches, feature engineering, and tuning the final model for every region. The final framework is called the Smooth Pinball Hybrid Neural Network (SPinHy-NN). The results showcased a suitable CNN architecture, providing more accurate forecasts compared to MLP and k-means clustering. The quantile cross-over problem has been evaluated through the crossing loss and the number of crossings metrics. The analysis shows that a margin parameter can correct this undesired behavior. The conclusions validate the SPinHy-NN as a generalized framework for different climates through the final score achieved post-competition, managing to capture the spatial patterns in most cases. Moreover, the framework can be adapted to reach a desirable trade-off between accuracy, sharpness, and consistency. Recommendations aim to extend this research by employing satellite imaging, further feature engineering, novel neural networks, and exploring spatiotemporal models to capture atmospheric dynamics.

During the preparation for the Olympic Sailing Competition, held in 2021 in Tokyo, Japan, the Dutch National Sailing Team encountered days with unpredicted wind behaviour. To gain more understanding in the wind patterns occurring, a deep learning based approach is taken. The goal of this research is to find out if unsupervised learning methods can contribute to wind pattern classification. It can then be investigated if the classification can increase understanding in specific wind patterns. The input data for the unsupervised learning model consists of 40 years of reanalysis wind speed data of an area including Japan. To classify the wind patterns, the dimensionality of the input data is reduced using different autoencoders. This reduced dimensional form is then clustered using K-means clustering. The results of the K-means algorithm are compared and the best autoencoder is chosen. The resulting clusters are analyzed for extreme wind patterns, such as typhoons. It is expected that these wind patterns will be clustered together. To check this, the cluster containing typhoon Jebi, the typhoon which caused the highest insurance cost ever in Japan, is analyzed. If this cluster contains typhoons, unsupervised learning is able to provide useful information regarding wind patterns. The best working autoencoder used in this research is the 3D CNN autoencoder. Using the 3D CNN autoencoder, some clusters with specific wind patterns are found. The cluster containing typhoon Jebi consists of 95.8% of typhoons, from which it can be concluded that unsupervised learning is a valid method for wind pattern classification.

Accurate short term rain predictions are important for flood early warning systems, emergency services, energy management and other services that that make weather dependent decisions. Recently introduced machine learning models suffer from blurry and unrealistic predictions at longer lead times, causing poor performance on the rarer heavy rainfall events. The objective of this research was to explore how the loss function in a recurrent, convolutional neural network (TrajGRU, X. Shi et al. (2017)) can be modified, to get sharper and more realistic predictions, without worsening its performance. Six perceptive loss functions, which should better represent how an image is perceived, are thoroughly analysed to understand the reasons behind their functioning in the context of precipitation nowcasting. It was shown that these perceptive losses can lead to an improvement in sharpness for the first lead time, but that the blurriness for longer lead times arises due to the imbalance in the dataset and the uncertainty of the model with respect to the location of the rain. This thesis gets concluded with recommendations on how to deal with these two problems.

Offshore wind energy is a renewable energy source that is anticipated to be significant in the global shift towards clean and green energy in order to reduce the reliance on fossil fuels and mitigate climate change. In recent years, the overall output of offshore wind energy has increased, and as a result, investments in offshore wind power have been on the rise. Governments around the world, particularly those in Europe, are investing heavily in establishing and maintaining offshore wind farms, with a particular focus on the North Sea. This is because the North Sea has high wind speeds and shallow waters, making it an ideal location for offshore wind farms....@en

Wind energy is becoming an important renewable energy source. An increased number of offshore wind farms are constructed due to the relatively higher wind speeds. Besides, compared with the land, the ocean areas offer more empty space for the installation of wind turbines. In recent years, several governments in Europe have the plan to expand their countries’ wind farms over the North Sea area. With this surge in the development of offshore wind farms, extreme weather events over the sea pose threats to the installations. A waterspout is one of such phenomenon of concern. In this study, we simulated and characterized the atmospheric conditions associated with two waterspout events observed recently over the North Sea. These cases were selected from the European Severe Weather Database. Various types of observational data, including radiosondes, radar reflectivities, satellite imageries, lighting maps, and floating liar-based wind profiles, were utilized for detailed characterization. Atmospheric circulation patterns associated with waterspouts were deduced from surface-level and upper-air synoptic charts. A mesoscale model, called the Weather Research and Forecasting model, was used for simulations with a high spatial resolution of 1 km. We used five different parameterizations of varying complexities to quantify the sensitivity of the simulated results with respect to cloud microphysics. A number of meteorological variables and indices (e.g., thermodynamic indices, wind shear, vertical velocity, reflectivity) are extracted from the simulations and compared with the observational data. In general, our results are in agreement with the findings from previous studies. For instance, we have found that a double moment microphysics parameterization produces more realistic results in comparison with a single moment one. However, we have noticed that our simulated results fall outside the range specified by the so-called Szilagyi waterspout nomogram. This nomogram was initially proposed based on observational data from the Great Lakes region and is widely used by the operational meteorologists. Based on the results, updating this nomogram is needed with additional observational and simulated data from the North Sea region.