TimesFM is a pre-trained univariate time-series forecasting model published in early 2024 [1]. Put simply, this model is capable of forecasting a sequence of data that is ordered in time (like hourly temperature or monthly flu cases) without any additional information or training. In the model's paper [2], the authors report that TimesFM was able to outperform both classical and deep learning models on well-known time-series datasets.
The goal of this project is to test TimesFM's accuracy on electricity consumption data on the Finnish transmission grid. Specifically, I am benchmarking the model on hourly consumption data from Jan 2024 to Dec 2024 by forecasting up to 24 hours ahead using the timesfm Python package [3]. The data was obtained from the operator of the Finnish grid through their Open Data portal [4].
Important
For a detailed write-up of the results, feel free to read the accompanying blog post on my personal webpage.
If you'd be interested in the code, please take a look at notebook.ipynb in this repository.
For a quick review of the results, below is a summary of the forecasting scheme and a barplot of monthly mean absolute percentage error (MAPE) values comparing the actual electricity consumption vs the model's forecasts:
Data
- electricity consumption in Finland in MWh
- hourly frequency (resampled from 15-min by averaging)
- start 2024-01-01 00:00, end 2024-12-31 23:00 (excluding some missing values)
Forecasting scheme
- up to 24 hours ahead
- start at 00:00, generate predictions up to 23:00 each day, and move to the next day
Model
- TimesFM 200M with default parameters
The code-related part (notebook.ipynb) of this repository is MIT-licensed. The electricity consumption data is licensed under CC BY 4.0, see [4].
[1] https://research.google/blog/a-decoder-only-foundation-model-for-time-series-forecasting/
[2] https://arxiv.org/abs/2310.10688
[3] https://github.com/google-research/timesfm
[4] "Electricity consumption in Finland" by Fingrid is licensed under CC BY 4.0