An Asymmetric Loss With Anomaly Detection LSTM Framework For Power Consumption Prediction

Abstract

Accurate load forecasting plays a pivotal role in energy management
systems, particularly for the residential sector, where unexpected
spikes or drops in consumption can lead to power outages or
inefficient resource utilization. This project presents a novel
framework combining Long Short-Term Memory (LSTM) networks
with asymmetric loss functions and DBSCAN-based anomaly
detection. The approach is designed to minimize underpredictions of
power consumption, which are more critical than overpredictions as
they can lead to electricity shortages.
Using three distinct datasets from France, Germany, and Hungaryeach
consisting of hourly electricity consumption, weather attributes,
and calendar information-the framework incorporates seasonality
splitting to reflect temporal usage patterns. Anomalies are first
detected and removed using the DBSCAN clustering algorithm to
improve data quality. Next, LSTM models are trained using various
asymmetric loss functions to penalize underestimations more severely
than overestimations. Our results demonstr...
Accurate load forecasting plays a pivotal role in energy management
systems, particularly for the residential sector, where unexpected
spikes or drops in consumption can lead to power outages or
inefficient resource utilization. This project presents a novel
framework combining Long Short-Term Memory (LSTM) networks
with asymmetric loss functions and DBSCAN-based anomaly
detection. The approach is designed to minimize underpredictions of
power consumption, which are more critical than overpredictions as
they can lead to electricity shortages.
Using three distinct datasets from France, Germany, and Hungaryeach
consisting of hourly electricity consumption, weather attributes,
and calendar information-the framework incorporates seasonality
splitting to reflect temporal usage patterns. Anomalies are first
detected and removed using the DBSCAN clustering algorithm to
improve data quality. Next, LSTM models are trained using various
asymmetric loss functions to penalize underestimations more severely
than overestimations. Our results demonstr...