Assignment 01: TinyML-Based Forecasting of Load, Renewable Energy Generation, and Energy Prices on Edge Devices
Summary
Develop and deploy ultra-lightweight forecasting models on the EnergyWatch edge device to predict short-term household electricity demand, local renewable generation, and electricity prices under strict embedded-computing constraints.
Problem Definition
Effective edge-based energy management requires reliable forecasts of electricity demand, renewable generation, and energy prices. These signals are strongly interdependent: household load is influenced by user behaviour and weather, renewable generation depends on local weather and irradiance conditions, and electricity prices reflect wider demand, supply, and renewable availability. However, most forecasting approaches are developed for cloud environments and are too computationally demanding for embedded edge devices. The challenge is to develop forecasting models that are accurate enough to support real-time control decisions while remaining small, fast, robust, and privacy-preserving. Within EnergyWatch, this means deploying forecasting intelligence directly on the edge device, close to the smart meter, where high-frequency data can be processed locally without unnecessary communication of household-level telemetry.
Method
The student will formulate a multi-horizon forecasting problem for household load, renewable generation, and electricity prices, using smart meter telemetry, time features, historical measurements, weather information, and price signals where available. The work will focus on developing TinyML-compatible models such as quantised recurrent neural networks, temporal convolutional networks, or compact attention-based models that can be deployed on the EnergyWatch edge device. A multi-task learning structure may be used, where a shared lightweight model backbone captures common temporal patterns while separate output layers generate forecasts for load, renewable production, and price.
The student will investigate model compression techniques such as quantisation, pruning, and knowledge distillation to make the models suitable for embedded deployment. The trained models will be converted to an edge-compatible runtime, such as TensorFlow Lite Micro or an equivalent framework, and tested directly on representative EnergyWatch gateway hardware. For renewable generation forecasting, the student may combine data-driven methods with simple physics-informed PV models to improve robustness under changing weather conditions. For price forecasting, the student will consider dynamic tariffs, day-ahead prices, or simplified price signals depending on data availability. The final models will be benchmarked against persistence models, statistical baselines, and independent single-task forecasting models. Evaluation will consider forecast accuracy, memory footprint, inference latency, energy consumption, deployment feasibility, and robustness to missing or noisy data.
Research Objective
The objective is to demonstrate that TinyML can provide reliable local forecasts for the main signals required by an edge-based energy management controller and that these models can be practically deployed on low-cost EnergyWatch edge devices. The student should show whether joint forecasting of load, renewable generation, and price provides better control-relevant information than separate forecasting models, while satisfying embedded hardware constraints. Success will be evaluated through forecasting accuracy, computational efficiency, robustness, real-device deployment performance, and the delivery of a deployable forecasting module integrated into the EnergyWatch edge tier.
Student Profile
A strong background in machine learning, embedded systems, and energy systems is required. Recommended courses include machine learning, deep learning, time series analysis, optimisation and control, embedded AI deployment, and embedded or IoT systems. Familiarity with Python-based ML frameworks and basic signal processing is expected.
For any questions please contact Eli Shirazi, e.shirazi@utwente.nl