Improving battery distribution
The ANWB has about 33 battery depots throughout the Netherlands where they supply battery’s to customers. They currently have no automated system which can predict when and how much to restock these battery depots. A team member of Votexa (Ruud) is responsible for making these predictions purely based on experience. This however starts to become more difficult now as Votexa is distributing new types of battery’s which don’t have a track record from the past.
I was tasked with developing a “load balancing algorithm” which could predict and forecast how many batteries would be needed at each battery depot for every 5 hours of the day (as Votexa restocks 3 times a day every 5 hours). They currently had no way to know how many batteries were present at each depot, as the distributers only record this once a day, which meant I had to figure out how Votexa could record this data in real-time. After some research, I made a report specifying how Votexa could make it easier for their employees working at these depots to record and publish changes (whenever they take a battery or return one) using a mobile application.
After this change in their process pipeline has been made, I could start analyzing the newly obtained data alongside scraping external internet sources to make a prediction of how many batteries at what locations they would sell at a specific moment in time. I had experiments with different time series and regression models to predict total sales per region, city and province. To make the model run with a higher throughput (so updating faster), I have optimized and parallelized the code using CUDA.
Based on these predictions, Ruud could make better judgements in what areas to restock and what the most optimal route would be to do this (based on distance and petrol used).