Gap filling and forecasting of high-frequency radar time-series data

Abstract

The real-time monitoring of the coastal and marine environment is vital for many reasons including oil spill detection and maritime security amongst others. The growing need for accurate real-time oceanographic observations and forecast data has lead to systems being set up on land and at sea to measure sea surface currents, sea level wave height and tides amongst others. Unfortunately, these networks can suffer from malfunctions caused by extreme weather conditions or network failure, which leads to a degradation in the monitoring system coverage, creating problems for the services making use of these real-time monitoring systems. Such degradation is subject to sporadic gaps in the High Frequency Radar (HFR) system which require to be filled accurately. To counter this problem, the use of machine learning architectures such as Feedforward neural networks, Recurrent neural networks and Random Forest models, have been investigated to perform gap-filling of the HFR data. Hyper-parameter tuning of these architectures was carried out to find the best configurations that fit the data effectively. Apart from the HFR data, satellite wind data was also considered to try enhance the prediction accuracy of these model architectures. Furthermore, look-back values between 3 and 24 hours were investigated to uncover the minimal amount of historical data required to make accurate predictions. Finally, drift in the data was also investigated to determine how often these model architectures might require re-training to keep them valid for predicting future data. Short-term forecasting was also investigated as an adaptation of the gap-filling model architecture. The neural network architectures were adapted to achieve forecasts of up to 24 hours. This amount of look-ahead is related to the data being recorded on a daily basis and being susceptible to changes over time. Look-back historical data was also considered for the purpose of short-term forecasting. The Long Short-Term Memory (LSTM) model architecture was found to be superior for gap-filling when compared to other models. The addition of wind data did not improve the prediction results further and a look-back of 6 hours was found to be the ideal minimal amount of historical data to use. Finally, no drift was found in the data and training the models on two years of data gave satisfactory results. For short-term forecasting, the LSTM model was found to produce the best forecasts over-all, using a look-back of 12 hours. A look-ahead of 24 hours was found to be adequate as extending the amount of look-forward further would lead to degradation in the forecasting results.