In this paper, probabilistic wind speed forecasts are constructed based on ensemble numerical weather prediction (NWP) forecasts for both wind speed and wind direction. Including other NWP variables in addition to the one subject to forecasting is common for statistical calibration of deterministic forecasts. However, this practice is rarely seen for ensemble forecasts, probably because of a lack of methods. A Bayesian modeling approach (BMA) is adopted, and a flexible model class based on splines is introduced for the mean model. The spline model allows both wind speed and wind direction to be included nonlinearly. The proposed methodology is tested for forecasting hourly maximum 10-min wind speeds based on ensemble forecasts from the European Centre for Medium-Range Weather Forecasts at 204 locations in Norway for lead times from 112 to 1108 h. An improvement in the continuous ranked probability score is seen for approximately 85% of the locations using the proposed method compared to standard BMA based on only wind speed forecasts. For moderate-to-strong wind the improvement is substantial, while for low wind speeds there is generally less or no improvement. On average, the improvement is 5%. The proposed methodology can be extended to include more NWP variables in the calibration and can also be applied to other variables.

Many data related problems involve handling multiple data streams of different types at the same time. These problems are both complex and challenging, and researchers often end up using only one modality or combining them via a late fusion based approach. To tackle this challenge, we develop and investigate the usefulness of a novel deep learning method called tower networks. This method is able to learn from multiple input data sources at once. We apply the tower network to the problem of short-term temperature forecasting. First, we compare our method to a number of meteorological baselines and simple statistical approaches. Further, we compare the tower network with two core network architectures that are often used, namely the convolutional neural network (CNN) and convolutional long short-term memory (convLSTM). The methods are compared for the task of weather forecasting performance, and the deep learning methods are also compared in terms of memory usage and training time. The tower network performs well in comparison both with the meteorological baselines, and with the other core architectures. Compared with the state-of-the-art operational Norwegian weather forecasting service, yr.no, the tower network has an overall 11% smaller root mean squared forecasting error. For the core architectures, the tower network documents competitive performance and proofs to be more robust compared to CNN and convLSTM models.

In this paper, we present the tower network, a novel, computationally lightweight deep neural network for multimodal data analytics and video prediction. The tower network is especially useful when it comes to combining different types of input data, a problem not greatly explored within deep learning.The architecture is further applied to a real-world example, where information from historic meteorological observations and numerical weather predictions are combined to produce high-quality forecasts of temperature for 1 to 6 hours into the future.The performance of the proposed model is assessed in terms of root mean squared error (RMSE), and the tower network outperforms even state-of-the-art forecasts from the Norwegian weather forecasting app yr.no from 3 hours into the future. On average, the RMSE of the tower network is approximately 6 % smaller than that of yr.no, and approximately 27 % smaller than that of the raw numerical weather predictions. CCS CONCEPTS• Applied computing → Earth and atmospheric sciences; • Computing methodologies → Machine learning; Computer vision.