Severe Weather Impact Database and Impact-based Forecasts by Utilizing Machine-Learning Technology

The availability of various digitized data sources has been steeply increasing during the last 15 years. However, in Finland next to no efforts have been made to gather and normalize weather-related impact data into a single data storage. To fill this gap of information a two-part project called SILVA was carried out in 2020-2022 to scan, collect and store weather-related impact data. In all, a dozen different data sources were identified ranging from electricity supply disruptions and road traffic accidents to airport traffic and train traffic flow anomalies. Most data series covered the whole country and ranged several years back in time. All datasets included time, location, and magnitude or strength of the impact events.

Once the impact events had been stored into the SILVA database they were utilized in various ways. First, some real-time impact data were combined with conventional meteorological data to offer quick impact views for nowcasting purposes (electricity supply disruptions, observed wind gusts and lightning observations etc.). Second, archived impact data and historical ECMWF weather forecasts were used for the training of gradient boosting machine-learning methodology. As a result, seven-day impact-based forecasts for the amount of damage clearance tasks, wildfire fighting tasks, traffic accidents and pedestrian slipping accidents were generated for each Finnish administrative region.

During the latter half of the project a seven-month pilot phase was carried out. More than 30 organizations attended the pilot, and they were provided with real-time weather observations, weather warnings, severe weather outlooks as well as novel impact-based forecasts. During the pilot phase severe weather forecasters gave video briefings ahead of each potentially interesting weather episode to make the users familiar with the pilot products. Briefings were held ten times, prior to blizzard, severe thunderstorm, wildfire, and synoptic-scale windstorm cases.

Preliminary verification results and the feedback gathered from the pilot users have been encouraging. The impact-based forecasts have been warmly welcomed by the users although some expected forecast pitfalls have been reported especially in relation to severe thunderstorm impacts. The users gave an overall rating of 4/5 on the project and rating of 4.3/5 on the future development potential.

Product example: 5-day impact-based forecast on the amount of wind damage clearance tasks for the Finnish administrative regions. Forecast issued on the 6th of August 2022 ahead of an severe thunderstorm event.