Wave Height Forecaster

Goal

To accurately predict the height of the waves in the vicinity of salmon farm pens.

Preamble

A particular salmon farming company operates its hatcheries in pens located offshore in mainly South-Eastern Tasmanian waters. Due to their location, they are often subject to wild weather which impacts the ability of the farmers to safely travel to them and complete the required maintenance tasks. These tasks include repairing the nets, feeding the fish and cleaning out waste. The ability to predict localised wave activity in advance would facilitate improved safety for the workers as well as better planning around scheduling the required tasks.

Data

A series of buoys located around the vicinity of the pens were equipped with accelerometers attached to radio-transmitters. These transmitters transmitted buoy movements in x, y and z axes at 5-minute intervals which were then stored in an Azure Cloud SQL instance in tabular format. The localised weather forecast was also derived from IBM’s The Weather Company and stored against the same 5-minute timestamp in the database. This importantly included wind strength and direction predictions, which were both found to have high attribute importance.


Method

The data was ingested into a Python script and loaded into a CNN LSTM model. This model was developed largely using a process of trial and error to determine the optimum hyper-parameters. This particular approach was used as prior research in this area had identified LSTM networks as being particularly good at time series data processing, as well as being ideal for extracting the features best used for the predictions.

Results

The idea was to be able to accurately predict the average wave heights 6-8 hours in advance, but our model was only useful for about an hour in advance. When looking at the literature, it became apparent that although LSTM networks are ideally suited for this task, a combination approach was best, where Random Forest combined with LSTM were most accurate.

The literature does support our findings that localised wave height predictions for significant periods of time in advance are difficult to achieve.

The model/approach definitely needs rethinking.

published August 23, 2020