A Unified Energy Management Framework for Multi-Timescale Forecasting in Smart Grids

TL;DR

Multi-pofo, a multi-scale power load forecasting framework, that captures such dependency via a novel architecture equipped with a temporal positional encoding layer is proposed and experimental results show that the approach outperforms compared to several strong baseline methods.

Abstract

Accurate forecasting of the electrical load, such as the magnitude and the timing of peak power, is crucial to successful power system management and implementation of smart grid strategies like demand response and peak shaving. In multi-time-scale optimization scheduling, rolling optimization is a common solution. However, rolling optimization needs to consider the coupling of different optimization objectives across time scales. It is challenging to accurately capture the mid- and long-term dependencies in time series data. This paper proposes Multi-pofo, a multi-scale power load forecasting framework, that captures such dependency via a novel architecture equipped with a temporal positional encoding layer. To validate the effectiveness of the proposed model, we conduct experiments on real-world electricity load data. The experimental results show that our approach outperforms compared to several strong baseline methods.

Figures (5)

• Figure 1: Building load as a superposition of different time scales, such as daily, weekly, monthly trends.
• Figure 2: Example of (a) campus level multi-timescale energy optimization architecture; (b) multi-timescale energy optimization process.
• Figure 3: Multi-pofo comprises three main components: a multi-scale embedding, a shared encoder-decoder and a multi-scale forecasting module. The model is trained in two stages: initially, the encoder and decoder are jointly trained using a reconstruction loss, followed by freezing the encoder and training the forecasting module with a prediction loss.
• Figure 4: Prediction error distribution across daily, weekly, and monthly scales for three different circuits (71, 193, and 272). Each boxplot represents the variability in prediction error (measured in kW), highlighting the mean (square markers), median (horizontal lines within the boxes), and range (whiskers). The daily scale shows larger variability and outliers compared to weekly and monthly scales, indicating the challenge of predicting short-term energy consumption. Weekly aggregation reduces variability, while monthly aggregation further smoothens prediction errors.
• Figure 5: Comparison of load forecasting between (a) Multi-pofo , and (b) CNN-LSTM, BiLSTM across day, week and month level.