Project Overview

Solar Power Forecasting Using PatchTST

I treated residential solar forecasting as a reliability problem: how can a home energy system make better decisions when sunlight, weather, and usage keep shifting?

January 2026 - June 2026

01Why It Matters

Problem

Residential solar power is useful only when people can plan around it. In practice, the output can drop suddenly because of clouds, rain, humidity, seasonal sunlight, panel temperature, or a mismatch between generation hours and household usage hours.

The real problem I wanted to address was not simply predicting a number. A homeowner needs an early sense of whether today is likely to be a high-production day, whether battery usage should be more careful, and whether grid electricity will probably be needed later.

Solar generation is noisy because weather variables and time-of-day patterns interact instead of changing independently.
Short-term forecast error can affect battery planning, grid draw, and when a household should run energy-heavy appliances.
A realistic forecasting project needs to expose error clearly, because a small-looking metric can still mean a noticeable operational mistake.

02What I Did

Decision

I framed the project as a reliability-focused time-series forecasting task and used PatchTST as the main reference architecture. The original PatchTST paper, accepted at ICLR 2023, proposes two central ideas: splitting time series into patch tokens and using channel independence for multivariate forecasting.

That design matched the solar setting because the model can look at longer historical context while keeping local temporal patterns intact. I treated the paper as a practical research anchor, then shaped my own pipeline around solar generation signals, meteorological variables, preprocessing, and repeatable evaluation.

Used the PatchTST idea of patch-level input tokens instead of treating every timestamp as an isolated token.
Connected solar generation data with weather-related features so the model was not learning from production history alone.
Evaluated the work with forecasting metrics such as MAE, RMSE, and MAPE so errors could be translated into practical reliability questions.

03What I Learned

Learning & Impact

The project became a foundation for my interest in renewable-energy intelligence and reliable AI for infrastructure-adjacent systems.

It also clarified how forecasting should be communicated. A user does not only need a curve; they need context, expected error, and a reason to trust or question the prediction.

Strengthened my understanding of transformer-based forecasting through an ICLR 2023 architecture rather than a generic model choice.
Connected machine learning experimentation with sustainability and household energy planning.
Created a technical base for future monitoring, anomaly detection, and predictive maintenance features.

Two people reviewing a laptop illustration

Takeaway

This project pushed me to think like both a researcher and a product engineer. PatchTST is interesting because the architecture is elegant, but for solar forecasting the important question is still practical: can the forecast help someone decide what to do before conditions change?

Research References

Nie et al., ICLR 2023

A Time Series is Worth 64 Words presents PatchTST as a channel-independent patch time-series Transformer for long-term forecasting and representation learning.

arXiv:2211.14730

The paper describes patching as a way to retain local semantics, reduce attention cost, and attend to longer history.

Stack & Links

PythonPyTorchPatchTSTPandasScikit-Learn

Back To Projects