Integrating text analysis in electricity load forecasting: initial findings from UK - Mines Paris Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

Integrating text analysis in electricity load forecasting: initial findings from UK

Intégration de l'analyse de texte dans la prévision de la charge électrique : premiers résultats au Royaume-Uni


The relationship between electricity load and weather has been established since long time and is one of the cornerstones in load prediction for operation and planning, along with behavioural and social aspects such as holidays or major events. This work presents the initial findings of research into the use of Natural Language Processing (NLP) for extracting previously unused information embedded into news or other forms of unstructured textual data for improving electricity demand forecasts. In this work, valuable sentiment and topic information is extracted from news texts with methods such as TextBlob and Latent Dirichlet Allocation (LDA) respectively. Multivariate time series are constructed by combining both text information and relevant numerical features (e.g. the number of news items per day). Granger causality test assists in finding which external features contribute to the forecasting of electricity load. At this point different time series forecasting approaches such as Autoregressive Integrated Moving Average Model, Support Vector Regression or LSTM are benchmarked. These methods are applied on a news dataset collecting from the BBC over the period between Jul 2015 and Dec 2021, and electricity load in the UK from the ENTSOe transparency platform. The results of this analysis are fed into a state-of-the-art load prediction model based on randomised decision trees. The performance over several horizons from days to months ahead is analysed according to a series of metrics for deterministic, probabilistic and ensemble forecasting. A comparison is made between two series of models, using or not using the textual features identified in the previous step. The objective is to understand under which conditions the proposed approach can improve the performance of the predictions.
ISFpresentation-YunBAI-FinalVersion.pdf (4.02 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03727324 , version 1 (19-07-2022)


  • HAL Id : hal-03727324 , version 1


Yun Bai, Andrea Michiorri, Simon Camal. Integrating text analysis in electricity load forecasting: initial findings from UK. 42nd International Symposium on Forecasting, Jul 2022, Oxford, United Kingdom. ⟨hal-03727324⟩
121 Consultations
65 Téléchargements


Gmail Facebook X LinkedIn More