Spatiotemporal Behind-the-Meter Load and PV Power Forecasting via Deep Graph Dictionary Learning

Area of Science:

• Electrical Engineering
• Data Science
• Renewable Energy Systems

Background:

• The proliferation of rooftop photovoltaic (PV) systems in distribution networks necessitates precise forecasting of behind-the-meter (BTM) load and PV generation.
• Current smart meter data, measuring only net load, limits the ability to quantify individual BTM load and PV generation, posing a challenge for grid operators.
• Accurate BTM measurements are essential for effective power system operation and grid stability with increasing renewable energy integration.

Purpose of the Study:

• To address the challenge of forecasting unobservable BTM load and PV generation by introducing the spatiotemporal BTM load and PV forecasting (ST-BTMLPVF) problem.
• To disaggregate historical net load data from neighboring residential units into their constituent BTM load and PV generation.
• To forecast future BTM load and PV generation time series for individual units.

Main Methods:

• Modeling residential units as a spatiotemporal graph (ST-graph), where nodes represent net load measurements and edges signify inter-unit correlations.
• Developing an ST-graph autoencoder (ST-GAE) to capture the complex spatiotemporal patterns within the ST-graph.
• Proposing a spatiotemporal graph dictionary learning (STGDL) optimization to extract salient spatiotemporal features from the ST-GAE's latent space for disaggregation and subsequent forecasting using a deep recurrent structure.

Main Results:

• The proposed STGDL method successfully estimates historical BTM load and PV generation by leveraging learned spatiotemporal features.
• The integrated deep recurrent structure accurately forecasts future BTM load and PV generation based on the disaggregated historical data.
• Numerical experiments on a real-world dataset demonstrate state-of-the-art performance for both BTM disaggregation and forecasting tasks.

Conclusions:

• The developed ST-BTMLPVF framework, utilizing ST-GAE and STGDL, effectively solves the problem of forecasting unobservable BTM load and PV generation.
• This approach provides a significant advancement over existing methods by enabling the quantification and prediction of individual BTM components.
• The findings offer a valuable tool for power system operators managing distribution networks with high rooftop PV penetration.