Explainable Load Forecasting with Covariate-Informed Time Series Foundation Models
This paper proposes a scalable SHAP-based explanation framework for Time Series Foundation Models, demonstrating that they can achieve competitive day-ahead load forecasting accuracy while maintaining the transparency and domain-aligned interpretability required for critical energy infrastructure.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to predict how much electricity a whole region will use tomorrow. In the past, experts used simple rules or statistical tricks. Now, we have "Time Series Foundation Models" (TSFMs). Think of these as super-smart, pre-trained AI chefs. They have tasted millions of different "recipes" (time series data) from all over the world. Because they've seen so much, they can cook up a forecast for a new kitchen (like a specific power grid) without ever needing to practice on that specific kitchen first. This is called "zero-shot" learning.
However, there's a problem. These AI chefs are black boxes. They give you a delicious prediction (the forecast), but they won't tell you why they added salt instead of sugar. In the power grid, where mistakes can cause blackouts, operators need to know why the AI thinks demand will be high. They need transparency, not just a number.
This paper introduces a new way to peek inside the chef's mind to see exactly which ingredients influenced the final dish.
The Problem: The "Black Box" Kitchen
The authors used two specific AI chefs: Chronos-2 and TabPFN-TS.
- Chronos-2 is like a chef who specializes in reading the rhythm of time (like a musician).
- TabPFN-TS is like a chef who is great at looking at a spreadsheet of facts and finding patterns.
Both are incredibly accurate at predicting electricity load. But if you ask them, "Why did you predict high usage for next Tuesday?" they usually just say, "Because my neural network said so." This is risky for critical infrastructure.
The Solution: The "Missing Ingredient" Test
To fix this, the authors created a method to calculate SHAP values. In simple terms, SHAP is a way to fairly divide the credit for a prediction among all the ingredients used.
Imagine you are baking a cake and it turns out delicious. To figure out how much credit the eggs deserve, you bake the cake again, but this time you leave out the eggs.
- If the cake falls apart, the eggs were crucial.
- If the cake tastes the same, the eggs didn't matter much.
The authors did this for electricity forecasting, but with a clever twist to make it fast:
- Time Masking: They asked the AI, "What if we pretend we don't know what the electricity usage was last week?" or "What if we pretend we don't know what happened yesterday?"
- Covariate Masking: They asked, "What if we pretend we don't know the weather forecast?" or "What if we pretend we don't know if it's a holiday?"
By systematically removing these "ingredients" (time steps or weather data) and seeing how the prediction changes, they could mathematically calculate exactly how much each factor contributed to the final answer.
The Results: The AI is Listening
The team tested these models on real electricity data from Baden-Württemberg, Germany. Here is what they found:
1. The AI is as good as the experts.
Even without being trained specifically on German data, these pre-trained models performed almost as well as models that had spent years studying German electricity data. They were also much faster to set up because they didn't need to be retrained.
2. The AI understands the world.
When the authors looked at the "explanations" (the SHAP values), the AI's reasoning matched human common sense perfectly:
- Holidays: When the model predicted lower usage on a Sunday or a public holiday, the explanation showed the "Holiday" ingredient was the main reason. It knew that factories close and people stay home.
- Temperature: On cold days, the model predicted higher usage. The explanation showed that the "Temperature" ingredient was driving this up, likely because people are using electric heating.
- Sunlight: When the sun was less bright than the day before, the model predicted higher grid usage. The explanation showed this was because solar panels (which usually help reduce grid load) were producing less power.
3. It's fast enough to use.
Usually, doing this "missing ingredient" test takes forever because there are so many combinations. But because these AI models are flexible (they can handle missing data easily), the authors could run the test in seconds. For example, explaining a forecast took about 5 seconds for Chronos-2.
The Bottom Line
This paper proves that we don't have to choose between accuracy and transparency. We can use these powerful, pre-trained AI models to predict electricity needs, and we can also ask them, "Why?" and get a clear, logical answer that matches our real-world understanding of weather, holidays, and human behavior.
It's like having a super-smart assistant who not only gives you the right answer but also hands you the recipe book and points to the exact page where they decided to add the extra sugar. This makes it safe and trustworthy to use these AI tools for keeping the lights on.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.