Want to increase forecasting accuracy 10%? Add a weather variable
We’re told weather data can improve the performance of predictive algorithms, help us understand risk and lead to useful insights. We decided to put these claims to the test.
In this post we build a simple probabilistic model to forecast UK energy demand. It turns out that by adding a single weather variable, predictive accuracy improves by 10%. Particularly at risk sensitive points in time, i.e. peaks in energy demand. We also see how uncertainty in our predictions changes, and drives useful insights.
With Bytehub’s simple-to-use data pipelines we grabbed 10 years of land surface temperature data from an accurate global weather model, 10 years of UK electricity demand and put them through our predictive model.
Below we plot actual demand against predicted demand with and without temperature.
Without temperature, the model has captured the main trends. With temperature, the model has captured the main trends but also seems to fit more of the variance. The improvement in accuracy turns out to be 10%!
Lets zoom in and see what is going on!
As suspected, without temperature the model captures the general weekly and monthly trends, but doesn’t pick up local fluctuations. With temperature, the model begins to predict local temperature related spikes and dives. Spikes and dives in demand equal risk for energy companies, networks and markets. By adding a relevant weather variable we have started to see this risk coming.
These observations aren’t restricted to energy: data from accurate weather models can critically enhance the performance of applications across governments and industries such as those involving aviation, automotive, construction, insurance, healthcare and supply chains to name a few.
At Bytehub AI we make it simple for anyone, data scientists, decision makers, analysts and AI experts, to pull in the right weather variable at the right location to predict risk, boost the performance of AI applications, and drive better decisions through visual analytics.
To find out more, visit
bytehub.ai/weather-for-data-science
or
