Short-Term Load Forecasting in a microgrid environment: Investigating the series-specific and cross-learning forecasting methods

Research output: Unpublished contribution to conferencePoster


A reliable and accurate load forecasting method is key to successful operation and energy management within new emerging smart grids in a microgrid environment. Due to the non-linear relations in data generating process, load forecasting remains a challenging task. While being well-established on an aggregate level, there is more uncertainty on a finer-grained scale of individual buildings. The data availability issue poses an additional challenge. An incremental roll-out of smart meters enables wider application of data-driven techniques among the researchers and in the industry. However, the individual time series may differ in terms of recorded temporal windows, temporal resolution and quality. Some assets may carry little data in order to successfully extract the underlying patterns of the series, such as seasonality and trend.
In this paper we investigate the application of deep learning in the short-term load forecast. The forecaster design is configured for a three days-ahead prediction. We focus on two model architectures: an artificial neural network (ANN) and a recurrent neural network (RNN). A variant of the ANN model with an additional feature encoding stage is investigated. The first stage generates forecasts for peaks and valley values alone and feeds them as inputs to the main predictor. For the RNN, we study the stacked long short-term memory models along with architectural variants. The potential accuracy gains from using weather data as additional input are investigated. The discussed preprocessing techniques include techniques in temporal feature engineering, normalization, addressing missing values, trend normalization, stabilizing the variance and deseasonalisation. An optimal strategy for preprocessing data is selected for the model in use. Additionally, the effect of cross-learning is investigated. Cross-learning is defined as a machine learning method where model utilizes learning from multiple series in order to predict individual one. The potential of cross-learning is studied by adapting the original recurrent neural network to a more global learning approach. Their accuracy is evaluated against a model that is trained individually for each series.
The effectiveness of proposed method is evaluated against established statistical techniques, such as autoregressive integrated moving average (ARIMA) and exponential smoothing (ETS) using accuracy indicators such as the root mean square error, the mean absolute percentage error, the mean absolute error and the Pearson correlation. The test datasets are multi-annual time series of electricity demand of buildings in hourly and halfhourly resolutions from open access sources (smart meter consumption in British Columbia, Canada and London, England) and private measurements in 15 minutes resolution of office buildings at the green energy park, a local energy community in in Zellik, Belgium. The proposed method of forecasting shows a significant improvement in comparison to existing statistical benchmarks. The examination of performance with cross-learning methods show potential for application in microgrid environment. The added benefits address limitations associated with building machine learning pipelines as seen by industry, such as scalability and latency.
Original languageEnglish
Publication statusPublished - 20 Mar 2023
EventMOBI Seminar 2023 - VUB, Brussels, Belgium
Duration: 31 Mar 202331 Mar 2023


SeminarMOBI Seminar 2023


Dive into the research topics of 'Short-Term Load Forecasting in a microgrid environment: Investigating the series-specific and cross-learning forecasting methods'. Together they form a unique fingerprint.

Cite this