PowerOracle

Inspiration

Our project was inspired by the chaos of last year's first pre-monsoon thunderstorm in Delhi. It was a kind of sticky heat where the ceiling fan just pushes warm air around, and the city's power demand was already high. When the storm hit suddenly, the load surged unpredictably as offices turned on lights and residential AC units worked overtime against the humidity.

A friend of ours working in grid operations mentioned how their forecasting models, which had predicted a normal Tuesday, were completely wrong-footed. They saw a massive, unexpected spike but couldn't immediately pinpoint the cause or anticipate its severity. At first, it just felt like a fluke one bad Tuesday. But the more we talked, the more we saw a bigger pattern: the models weren’t built for surprises. the problem: existing models are good at predicting the expected, but fail when reality strikes. We wanted to build a tool that could see the storm coming both literally and figuratively and explain its impact.

What it does

Power Oracle is an intelligent forecasting system that tells grid operators the what, when, and why of energy demand.

It Forecasts: It provides highly accurate, short-term energy load predictions using a state-of-the-art Time Series Foundation Model.
It Enriches: Unlike other tools, it incorporates real-world context like real-time weather forecasts, public holidays, and major local events (like a cricket match) to make its predictions smarter.
It Explains: When an anomaly is detected, Power Oracle doesn't just raise an alarm. It uses Explainable AI (XAI) to instantly diagnose the cause, automatically labeling it as "Weather-Driven," "Atypical Holiday Pattern," or an "Unexplained System Anomaly" that needs urgent attention. This is all presented in a simple, actionable dashboard.

How we built it

We built Power Oracle as a modular, end-to-end system.

Core Forecasting Engine: We leveraged a pre-trained Lag-Llama TSFM for its powerful zero-shot forecasting capabilities.
Context Engine: A Python script using Pandas pulls and processes data from a public Weather API and a custom calendar of local holidays and events.
Explainability Layer: We integrated the SHAP library to analyze the model's predictions. When an anomaly is flagged, a SHAP analysis runs to attribute the deviation to specific input features.
Frontend Dashboard: A user-friendly interface built with Streamlit visualizes the forecast, logs anomalies, and displays the simple, auto-generated natural language explanations.

Challenges we ran into

Our biggest challenge was data homogenization. The TSFM expects clean numbers, but our contextual data was a mix of numerical weather data, categorical holidays, and unstructured text from news events. Creating a pipeline to transform all these into a unified format that the model could understand was a complex but rewarding task.

Another hurdle was making the XAI outputs genuinely useful. A SHAP waterfall plot is great for a data scientist, but not for a busy grid operator. We spent a lot of time designing logic to automatically translate these complex charts into simple, one-sentence summaries.

Accomplishments that we're proud of

We're incredibly proud of moving beyond simple anomaly detection to creating true anomaly intelligence. Our system doesn't just say something is wrong; it tells you why, which is the missing piece in today's grid management tools.

Building a fully functional, end-to-end prototype—from live data ingestion to an interactive, explainable UI—within the hackathon's timeframe is an accomplishment our whole team is proud of. We successfully closed the "trust gap" between the human operator and the AI.

What we learned

Context is King: We learned that for real-world forecasting, what's happening outside the data (like the weather or a festival) is just as important as the historical data itself.
The Last Mile of AI is User Experience: A powerful model is useless if its insights aren't accessible. The design of the dashboard and the clarity of the explanations were just as critical as the model's accuracy.
Leverage, Don't Rebuild: Using a pre-trained foundation model was a game-changer. It allowed us to focus our energy on the innovative parts of our project—the context engine and the XAI layer—instead of spending all our time training a model from scratch.

What's next for PowerOracle

The future for Power Oracle is to evolve from a reactive monitoring tool into a proactive decision-support system. Our next steps include:

Building a "What-If" Scenario Planner: Allow operators to ask questions like, "What will the load look like if tomorrow's heatwave is 2°C hotter than forecasted?"
Integrating More Real-Time Data: Incorporate live traffic data and social media trends to detect unscheduled events that impact energy use.
Automated Recommendations: Evolve the system to not only explain anomalies but also to suggest optimal responses, like demand-response actions or grid reconfigurations.