TL;DR
This paper explores how to improve electricity load demand (ELD) forecasting by incorporating human behavior data. They found that combining traditional electricity consumption data (Low Voltage for residential, High Voltage for industrial) with Human Dynamics (HD) data, like people's movement and presence in an area, significantly reduces prediction errors, especially in larger, more dynamic regions.
What Problem Does This Paper Address?
Accurately predicting electricity load demand (ELD) is crucial for grid stability and efficient energy management. ELD is heavily influenced by human behavior, which manifests in daily routines and mobility patterns. Traditional forecasting models often struggle to capture these nuances, especially when dealing with the complex spatiotemporal (space and time) nature of electricity consumption. The COVID-19 pandemic highlighted this, showing how shifts in human behavior (like lockdowns and remote work) drastically altered consumption patterns. This paper aims to bridge this gap by investigating how different types of data representing human activity—specifically, High Voltage (HV) electricity consumption (proxy for industrial/commercial activity) and Human Dynamics (HD) data (location-based mobility)—can enhance Low Voltage (LV) consumption forecasts at a granular, mesh-level.
Methods
The researchers employed a deep learning approach, specifically using the Informer model, a Transformer-based architecture known for its efficiency in long-sequence time-series forecasting. They framed the problem as a multivariate spatiotemporal forecasting task. The core of their methodology involved experimenting with various combinations of input features:
- LV (Low Voltage) data: Represents residential and small commercial electricity consumption.
- HV (High Voltage) data: Represents industrial and large commercial/office electricity consumption.
- HDM (Human Dynamics Move): Counts of people moving within a mesh.
- HDS (Human Dynamics Stay): Counts of people staying within a mesh.
They evaluated eight different feature combinations across three progressively larger geographical "sets" of meshes (U, V, and W) to understand the impact of spatial coverage and feature availability. The primary metric for evaluation was the spatiotemporal prediction error (STPE), with spatiotemporal standard deviation (STSD) used to assess the consistency of improvements.
Datasets
The study utilized data from Utsunomiya City, Japan, divided into 500m square meshes.
- Electricity Load Demand (ELD): Hourly LV and HV consumption data per mesh from January 2017 to December 2020, sourced from GridDataBank Lab (GDBL).
- Human Dynamics (HD): Half-hourly counts of people staying (HDS) and moving (HDM) within meshes from October 2017 to October 2020, provided by KDDI Corporation (anonymized and statistically processed smartphone location data).
The study period for training, validation, and testing was carefully selected to avoid COVID-19 pandemic anomalies, focusing on data from October 2017 to March 2020.
Key Insights and Ideas
- HV data is a strong proxy for commercial/industrial activity: Even when not directly forecasting HV, its patterns provide valuable context for understanding broader economic activity that influences overall electricity demand.
- Human Dynamics (HD) data shines in dynamic, widespread areas: HD data is particularly effective when there's significant human movement and presence across a large geographical area, capturing the ebb and flow of daily life.
- The "best" features depend on the area: There's no one-size-fits-all solution. For dense urban areas with strong business activity, HV data might be more consistently beneficial. For sprawling regions with diverse human activity, HD data becomes more powerful.
- Adding features can increase variability: While adding more data generally reduces average prediction error (STPE), it can sometimes lead to more inconsistent improvements across different areas (higher STSD). This suggests a trade-off between overall accuracy and uniform performance.
- Mesh-level analysis reveals spatial nuances: Visualizing prediction errors across meshes (Figure 2) shows that models can overfit to certain areas, achieving high improvements there while potentially degrading performance elsewhere. This highlights the importance of considering feature selection based on local characteristics.
Key Results
- HV + LV: Combining HV with LV data reduced STPE by an average of 13.37% and STSD by 46.7% compared to LV-only.
- HD + LV: Integrating HD data (HDM, HDS, or both) with LV data achieved an average STPE reduction of 14.3% and STSD reduction of 54.6% compared to LV-only.
- HV vs. HD: For the largest set (W), HD features improved STPE by 20.69% and STSD by 40.91% on average, outperforming HV. However, for smaller, denser sets (U), HV showed better relative improvements.
- Combined Features: Scenarios combining LV, HV, and HD data generally improved performance but didn't always universally lower the overall prediction error, sometimes increasing STSD.
- Best performing scenarios: For set W, using HDS achieved the best STPE. For set V, combining HV and HDS yielded the best STPE.
Limitations and Future Work
- Data Availability: Access to ELD data was limited to one month, constraining the scope of experimentation. HD data was also not publicly available.
- Geographical Scope: The study was limited to Utsunomiya City, which may limit the generalizability of findings to other cities with different characteristics.
- HD Data Limitations: HD data, while useful, doesn't capture the flow of people between meshes, which can be a limitation in smaller, more interconnected areas.
- Future Work: The authors suggest automating mesh labeling, refining the definition and application of HD data, and exploring advanced modeling techniques to better capture human behavior's influence on ELD. They also propose further experimentation to test generalizability across different cities.
Summary updated