Consumer Energy Demand Distribution Forecasting using Clustering Strategies and Applications in Robust Peak Shifting with Batteries
Summary
In recent years, the electric grid is becoming increasingly overloaded. At the same time, dependence on electric energy is growing, and production is becoming less plannable and predictable due to a growing share in renewable energies.
To alleviate the resulting peak stresses on the electric grid, we perform peak shifting using batteries at the neighborhood scale. The (dis)charging schedules are based on a self-made hourly consumer demand forecasting approach.
The LSTM architecture is proposed for forecasting, and compared to an LR approach. Instead of a point-forecast, we implement a distribution forecast, allowing for accurate modelling of the stochasticity of consumer demand. Clustering is used for efficient training, and improved by our ReClustering technique, in which consumers are reclassified based on forecasting performance.
For planning batteries, we use 5 or 10 Tesla Powerwall batteries for peak shifting the hourly demand of the 500 consumers at hand. We propose and prove optimality of a fully polynomial time approximation scheme for the offline case, and evaluate different strategies in the online approach, in which a rolling horizon approach is used to solve for a time horizon of 24 hours into the future. A variable battery reserve is included in our proposed approach.
Our LSTM forecasting technique outperforms the shift (a persistence model) and linear regression approaches. Clustering significantly improves results. ReClustering also contributes to improved results.
Our battery planning algorithm outperforms the baseline approaches that do not use our forecasting approach, but only when the information in the percentiles of the distribution forecast are used for deciding on battery reserve, showing one of the advantages of distribution forecasting over point forecasting.