[Paper Review] A Single Scalable LSTM Model for Short-Term Forecasting of Disaggregated Electricity Loads

Andrés M. Alonso, Francisco J. Nogales+1 more|arXiv (Cornell University)|Oct 15, 2019

Energy Load and Power Forecasting28 references4 citations

TL;DR

This paper proposes a single, scalable Long Short-Term Memory (LSTM) recurrent neural network model that forecasts short-term electricity loads for thousands of disaggregated smart meters simultaneously. By training on a representative subset of heterogeneous households, the model achieves over 20% better accuracy than univariate benchmarks, even forecasting new, unseen consumers with minimal computational cost after training.

ABSTRACT

Most electricity systems worldwide are deploying advanced metering infrastructures to collect relevant operational data. In particular, smart meters allow tracking electricity load consumption at a very disaggregated level and at high frequency rates. This data opens the possibility of developing new forecasting models with a potential positive impact in electricity systems. We present a general methodology that is able to process and forecast a large number of smart meter time series. Instead of using traditional and univariate approaches, we develop a single but complex recurrent neural-network model with long short-term memory that can capture individual consumption patterns and also consumptions from different households. The resulting model can accurately predict future loads (short-term) of individual consumers, even if these were not included in the original training set. This entails a great potential for large scale applications as once the single network is trained, accurate individual forecast for new consumers can be obtained at almost no computational cost. The proposed model is tested under a large set of numerical experiments by using a real-world dataset with thousands of disaggregated electricity consumption time series. Furthermore, we explore how geo-demographic segmentation of consumers may impact the forecasting accuracy of the model.

Motivation & Objective

• To develop a scalable, single-model approach for short-term forecasting of massive, disaggregated electricity load time series from smart meters.
• To overcome the computational burden of training individual models per household by using a unified LSTM architecture.
• To enable accurate forecasting of new, previously unseen consumers without retraining.
• To evaluate the impact of geo-demographic segmentation on forecasting performance.
• To provide a computationally efficient, large-scale solution for real-time load forecasting in smart grid applications.

Proposed method

• A single LSTM-based recurrent neural network is trained on a representative subset of smart meter time series grouped by geo-demographic characteristics.
• The model jointly learns patterns across multiple households, capturing individual consumption behaviors and inter-household dynamics.
• Input features include temporal variables (hour, day of week, month), weather data, and calendar effects to enhance forecasting accuracy.
• The model is trained once and reused for out-of-sample forecasting of new households, even those not in the training set.
• Performance is evaluated using MAE (Mean Absolute Error) across multiple out-of-sample periods and geo-demographic groups.
• A hierarchical approach is used: first, data is segmented into 19 homogeneous groups based on geo-demographic profiles; second, a single LSTM is trained per group, or globally on a mixed set for generalization.

Experimental results

Research questions

• RQ1 Can a single LSTM model achieve high forecasting accuracy across thousands of disaggregated electricity load time series?
• RQ2 How does the model perform when forecasting new, unseen households not included in the training set?
• RQ3 To what extent does geo-demographic segmentation improve forecasting accuracy compared to a global model?
• RQ4 What is the computational efficiency of the proposed model compared to traditional univariate approaches?
• RQ5 How does the model handle non-linearities, seasonality, and noise in individual household load data?

Key findings

• The proposed LSTM model outperforms univariate benchmarks by more than 20% in median MAE across all testing periods and meters in the back-testing experiment.
• The model achieves consistent out-of-sample performance across all 19 geo-demographic groups, including the ACORN-U and Global groups, with stable error reduction.
• Even when forecasting new households not in the training set, the model maintains high accuracy, demonstrating strong generalization capability.
• The model's performance is significantly enhanced when trained on homogeneous geo-demographic groups, but remains effective even without prior classification.
• The model achieves scalable, near real-time forecasting for hundreds of thousands of smart meters with only one initial training run.
• The success of the model is attributed to its ability to capture complex, non-linear patterns and seasonalities through joint learning across large datasets.