Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Electricity Demand Forecasting in Future Grid States: A Digital Twin-Based Simulation Study

Daniel R. Bayer, Felix Haag, Marco Pruckner, Konstantin Hopf

TL;DR

This study tackles short-term residential electricity demand forecasting under future grid states with decentral generation and sector coupling. It integrates a digital twin of a local energy system with day-ahead ML forecasts (notably LSTM and CNN-LSTM) and compares them to Synthesized Load Profiles and other benchmarks, using data from $3{,}511$ households over $34$ months. Results show ML methods can reduce RMSE by up to $68.5\%$ in the current grid state, but forecast errors rise as PV penetration increases in future states, highlighting the need for continued methodological advances and twin-enhanced planning. The work demonstrates the utility of combining digital twins with ML for utilities and grid operators, and points to Transformer-based approaches and richer twin models as promising directions for robust forecasting in future energy systems.

Abstract

Short-term forecasting of residential electricity demand is an important task for utilities. Yet, many small and medium-sized utilities still use simple forecasting approaches such as Synthesized Load Profiles, which treat residential households similarly and neither account for renewable energy installations nor novel large consumers (e.g., heat pumps, electric vehicles). The effectiveness of such "one-fits-all" approaches in future grid states--where decentral generation and sector coupling increases--are questionable. Our study challenges these forecasting practices and investigates whether Machine Learning (ML) approaches are suited to predict electricity demand in today's and in future grid states. We use real smart meter data from 3,511 households in Germany over 34 months. We extrapolate this data with future grid states (i.e., increased decentral generation and storage) based on a digital twin of a local energy system. Our results show that Long Short-Term Memory (LSTM) approaches outperform SLPs as well as simple benchmark estimators with up to 68.5% lower Root Mean Squared Error for a day-ahead forecast, especially in future grid states. Nevertheless, all prediction approaches perform worse in future grid states. Our findings therefore reinforce the need (a) for utilities and grid operators to employ ML approaches instead of traditional demand prediction methods in future grid states and (b) to prepare current ML methods for future grid states.

Electricity Demand Forecasting in Future Grid States: A Digital Twin-Based Simulation Study

TL;DR

This study tackles short-term residential electricity demand forecasting under future grid states with decentral generation and sector coupling. It integrates a digital twin of a local energy system with day-ahead ML forecasts (notably LSTM and CNN-LSTM) and compares them to Synthesized Load Profiles and other benchmarks, using data from households over months. Results show ML methods can reduce RMSE by up to in the current grid state, but forecast errors rise as PV penetration increases in future states, highlighting the need for continued methodological advances and twin-enhanced planning. The work demonstrates the utility of combining digital twins with ML for utilities and grid operators, and points to Transformer-based approaches and richer twin models as promising directions for robust forecasting in future energy systems.

Abstract

Short-term forecasting of residential electricity demand is an important task for utilities. Yet, many small and medium-sized utilities still use simple forecasting approaches such as Synthesized Load Profiles, which treat residential households similarly and neither account for renewable energy installations nor novel large consumers (e.g., heat pumps, electric vehicles). The effectiveness of such "one-fits-all" approaches in future grid states--where decentral generation and sector coupling increases--are questionable. Our study challenges these forecasting practices and investigates whether Machine Learning (ML) approaches are suited to predict electricity demand in today's and in future grid states. We use real smart meter data from 3,511 households in Germany over 34 months. We extrapolate this data with future grid states (i.e., increased decentral generation and storage) based on a digital twin of a local energy system. Our results show that Long Short-Term Memory (LSTM) approaches outperform SLPs as well as simple benchmark estimators with up to 68.5% lower Root Mean Squared Error for a day-ahead forecast, especially in future grid states. Nevertheless, all prediction approaches perform worse in future grid states. Our findings therefore reinforce the need (a) for utilities and grid operators to employ ML approaches instead of traditional demand prediction methods in future grid states and (b) to prepare current ML methods for future grid states.

Paper Structure

This paper contains 14 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Research approach
  4. Digital twin modeling and simulation of future grid states
  5. Dataset on current grid state
  6. PV expansion scenarios for modeling future grid states
  7. Simulation results for current and future grid states
  8. Predictive modeling
  9. Training and data preparation
  10. Evaluation metrics
  11. Benchmark predictors
  12. Results
  13. Discussion
  14. Conclusion and future work

Figures (3)

  • Figure 1: Experimental design.
  • Figure 2: Left: Violinplot of one-year electric demand on grid level caused by residential buildings in kW averaged over one hour for current grid state (CS) and the scenarios S1 and S2. Right: Same plot for electric load in kW. The negative values occur due to a strong feed-in from residential PV installations that cannot be consumed locally.
  • Figure 3: Forecasting results (RMSE) of ML models and naive estimators in a current (a) and future S1 (b) and S2 (c) grid state unfolded over all hours of the day. The higher errors during the day can be attributed to the uncertain actual PV production.