PaperlayerView exact revision on arXiv

Source-linked AI summary

News and Load: A Quantitative Exploration of Natural Language Processing Applications for Forecasting Day-ahead Electricity System Demand

Yun Bai, Simon Camal, Andrea Michiorri

arXiv:2301.07535v2cs.CLcs.AIcs.CY

TL;DR

This paper examines whether social information in news can help explain and forecast electricity demand, where prior work leaves the text-demand relationship and its mechanisms insufficiently explored. It integrates news-derived features with structured predictors and analyzes their relationship to demand. The authors report improvements across three forecast error metrics and identify a best-performing combination of textual features.

  • Problem

    The paper investigates whether social factors visible in news influence electricity demand and whether NLP can quantify text for forecasting.

    Source: §I · pp. 1–2

  • Method

    The authors convert news into numerical text features, filter them with a Granger test, and add them to a forecasting model with demand, calendar, and temperature predictors.

    Source: §V · p. 11

  • Results

    4% RMSE, 11% MAE, and 10% SMAPE improvements are observed; the best-performing method combines title word frequencies, sentiment scores, and GloVe embeddings from news bodies.

    Source: §V · p. 11

  • Takeaways & Limitations

    The experiments demonstrate the potential of textual features for short-term and regional electricity-demand forecasting.

    Source: §V · p. 11

    The reported improvements should be considered in light of the relatively smooth behavior and predictability of national aggregated load.

    Source: §V · p. 11

Abstract

from arXiv · show

The relationship between electricity demand and weather is well established in power systems, along with the importance of behavioral and social aspects such as holidays and significant events. This study explores the link between electricity demand and more nuanced information about social events. This is done using mature Natural Language Processing (NLP) and demand forecasting techniques. The results indicate that day-ahead forecasts are improved by textual features such as word frequencies, public sentiments, topic distributions, and word embeddings. The social events contained in these features include global pandemics, politics, international conflicts, transportation, etc. Causality effects and correlations are discussed to propose explanations for the mechanisms behind the links highlighted. This study is believed to bring a new perspective to traditional electricity demand analysis. It confirms the feasibility of improving forecasts from unstructured text, with potential consequences for sociology and economics.

I. INTRODUCTION

The paper explores whether news text captures social influences on electricity demand and can improve forecasting, addressing a gap in prior energy studies that used limited textual information or did not examine how text relates to forecasts. It combines news-derived features with structured predictors in a forecasting chain and analyzes their relationships with demand.

I. INTRODUCTION

  • News-derived information is tested as a way to improve electricity-demand forecasting and explain links between social events and demand.The motivation includes global events that can create unseen market conditions and challenge previous assumptions.Source: §I · p. 1
  • Prior energy studies considered social data or text, but did not discuss how textual information relates to forecasts or how its mechanisms work.Examples include work using tweet counts without semantic information and work relating crowdsourced data to consumption without analyzing text contents.Source: §I · pp. 1–2
  • The study tests whether social factors visible in news influence demand and whether NLP can quantify text for practical forecasting use.Its stated contributions include a complete chain combining text and structured data, plus analysis of performance mechanisms from global, local, and causality perspectives.Source: §I · p. 2

II. METHODOLOGY

  • The forecasting workflow acquires load, temperatures, calendar data, and news, then compares text-augmented models with benchmark and official forecasts.The benchmark uses demand lags, calendar features, and temperatures; the text features are filtered with a Granger test before evaluation.Source: §II · p. 2

A. Pre-processing for the text-based model

  • Text preprocessing creates word lists by tokenizing sentences, lowercasing words, and removing stopwords, words under three letters, and numbers.The steps are intended to remove irrelevant terms, aggregate contents, and increase semantic information.Source: A. Pre-processing for the text-based… · p. 2

1) Statistical methods:

The paper converts news into statistical, semantic, and representative text features for forecasting. These include daily counts and word frequencies, sentiment and topic distributions, and GloVe embeddings.

1) Statistical methods:

  • Count features summarize daily news volume and website-section proportions, while word-frequency features count selected words that pass frequency thresholds.The study uses 27 count series; it selects 456 title words, 329 description words, and 550 body words.Source: 1) Statistical methods · p. 2

2) Semantic methods:

  • Sentiment features include polarity and subjectivity summaries, while LDA topic distributions average each day's news-topic probabilities.Topic count is assessed by comparing Topic Coherence and Average Topic Overlap across κ values; the figure marks ideal κ values of 87 for T, 100 for D, and 69 for B.Source: 2) Semantic methods · p. 3, Figure 2 · p. 3

3) Representative methods:

  • GloVe maps words to 100-dimensional vectors, which are averaged within each news item and then across the day to form embedding features.The study uses pre-trained Stanford GloVe embeddings.Source: 3) Representative methods · p. 3

4) Granger test:

The Granger test filters textual features by testing their predictive contribution to demand, while day-ahead forecasts combine retained text with established predictors in an ExtraTrees model. The selected 30-lag tests had the lowest AIC, but the authors caution that this filtering is not a causality test.

4) Granger test:

  • The test checks whether a stationary textual-feature series contributes to forecasting demand and serves as a feature filter to reduce a large textual feature set.The paper distinguishes this filtering role from its later discussion of causality.Source: 4) Granger test · p. 3
  • The bilateral Granger procedure retained features when X predicted demand Y at p1 < 0.05 but demand did not predict X at p2 ≥ 0.05.Nonstationary series were differenced after an Augmented Dickey-Fuller test, and lag orders were chosen by AIC from 1, 7, 30, and 90 days.Source: 4) Granger test · p. 3
  • 30 lags produced the lowest AIC and selected fewer than 100 of more than 2,000 textual features; this filtering is not an actual causality test.The authors note that correlations between lagged and predicted values do not necessarily imply causality.Source: Table I · p. 4, 4) Granger test · p. 4

B. Forecasting with textual features

  • ExtraTrees forecasts use concatenated demand lags, calendar, temperature, and textual features to predict 48 half-hourly demand values; ensemble forecasts average tree predictions.Each tree prediction is determined by the feature-space region containing the input.Source: B. Forecasting with textual features · p. 4

C. Evaluation

The forecasting models are assessed with test-period averages of RMSE, MAE, and SMAPE. The paper examines model behavior through global correlations, local feature explanations, and a separate causality analysis.

C. Evaluation

  • RMSE, MAE, and SMAPE are calculated at each test time step and averaged over the whole test period.The forecasting horizon is 48 half-hourly steps.Source: C. Evaluation · p. 4

D. Explanation of the model

  • Global explainability uses Pearson correlations, local explainability uses LIME, and causality is analyzed with Double Machine Learning.The paper uses these analyses to explore mechanisms behind text-related forecasting improvements.Source: D. Explanation of the model · p. 4

1) Local explainability:

  • LIME perturbs a sample to create a local dataset, fits a simple surrogate such as linear regression, and uses its coefficients to interpret feature effects on that forecast.The local dataset is normally distributed and centered on the current sample.Source: 1) Local explainability · p. 4

2) Causality effects:

The study uses double machine learning to examine whether selected news features have effects on demand predictions, alongside a benchmark built to compare textual features with official forecasts. Its dataset pairing and benchmark results set the context for those analyses.

2) Causality effects:

  • Double ML is used to estimate a feature's effect on predictions while other features remain constant, because correlation alone does not establish causality.ExtraTrees regressors supply the nuisance-function forecasts, and OLS estimates the treatment effect.Source: 2) Causality effects · p. 4

E. Datasets

  • The authors paired electricity demand with BBC news, selecting data for shared geography and temporal overlap, authoritative news, accessible language, and open access.The five-year dataset covers June 2016 to May 2021; over 80,000 BBC news items were collected.Source: E. Datasets · pp. 4–5

A. Benchmark model

  • Combining demand lags, calendar features, and temperature performed close to ENTSO-E across three metrics; holidays added no improvement, while best gains were around 200MW in mae and 4% in smape.The authors built ExtraTrees as a similar-performing, reproducible alternative because ENTSO-E's method was not publicly available.Source: A. Benchmark model · p. 5, Table II · p. 5

B. Textual feature-enhanced model

The paper adds news-derived textual features to its benchmark and finds improvements from multiple feature groups, with word frequencies especially useful in titles.

B. Textual feature-enhanced model

  • The textual inputs comprise count features, word frequencies, sentiment, topic distributions, and GloVe word embeddings drawn from news titles, descriptions, and bodies.The study reports original and post-Granger-test selected feature counts, while Table IV describes features within each group.Source: 1) Impact of textual features · p. 5, Table IV · p. 6
  • Six groups of textual features improve the benchmark; word frequencies in titles, descriptions, and bodies reduce errors, especially in titles.Sentiment, topic distributions, and GloVe embeddings from news bodies also outperform those extracted from titles and descriptions.Source: 1) Impact of textual features · p. 5, Table III · p. 5
  • Improving word-frequency features include names, European places, political terms, and COVID-19-related words, while body-derived embeddings capture weapons, transportation, and military issues.The minimum subjectivity value is influential, and topics include Northern Ireland politics and coronavirus.Source: 1) Impact of textual features · p. 5

2) Feature combination:

M6 improves short-term day-ahead forecasting over the ExtraTrees benchmark, while feature combinations and model comparisons further characterize the results. Gains are more marked during early-morning and evening demand ramps.

2) Feature combination:

  • M6, combining body sentiment and GloVe embeddings with title word frequencies, has the best performance among the tested combinations and is used for further analysis.Adding body sentiment reduces errors, while body GloVe embeddings reduce their spread; adding description and body word frequencies to title word frequencies increases errors.Source: 2) Feature combination · p. 6, Figure 4 · p. 7

3) Error analysis and DM-test:

  • Forecasting improvements are generally more marked in the early morning and evening, when demand ramps are usually larger.Figure 5 compares errors with and without textual features across hours.Source: 2) Feature combination · p. 6, Figure 5 · p. 7
  • The DM test finds M0 significantly superior to D + C + T, with p=0.0404; the test treats p-values below 0.05 as significant.Table V reports pairwise model-comparison p-values.Source: 2) Feature combination · p. 6, Table V · p. 6

4) Results on short-term forecasting:

C. Textual feature variation and sensitivity analysis

The section tests how news volume and selection affect text-based forecasting, finding that using all available news generally outperforms prior-knowledge selection.

1) Variation in the amount of news:

  • The analysis studies the trade-off between retaining more news information and reducing training time by comparing all-news inputs with news selected by section.The authors aggregate feature groups by section for demand forecasting; the news-section statistics are presented in Table VII.Source: 1) Variation in the amount of news · p. 7, Table VII · p. 7
  • TD features outperform the benchmark from every section except Business, and NIE-section topics outperform topics from the news as a whole.The NIE topics correspond to themes including financial fraud, education, government response to Covid-19, vaccines, and immigration in Europe.Source: 1) Variation in the amount of news · p. 7
  • CF features from Europe, Wales, and Electricity improve the benchmark by less than 1%, while GWE features from Wales, London, and Electricity improve RMSE and MAE but have higher standard deviation than all-news results.Table VIII reports sectional-news forecasting errors; the prose identifies these comparisons.Source: Table VIII · p. 8, 1) Variation in the amount of news · p. 8
  • All available news can enhance forecasts, while news selected using prior knowledge may not help.The authors argue that all-news inputs retain broader social information and yield more accurate results, despite longer NLP training; deployment is feasible with current computing infrastructure.Source: 1) Variation in the amount of news · p. 8

2) Comparing different word-embedding methods:

The paper compares GloVe, Word2Vec, and TF-IDF embeddings, then tests textual features for Northern Ireland demand; GloVe leads on RMSE and MAE, while TF-IDF leads on SMAPE.

2) Comparing different word-embedding methods:

  • TF-IDF weights document-specific word importance without semantic relationships, Word2Vec models local context, and GloVe incorporates global semantic relationships.The paper uses 100-dimensional Word2Vec vectors, matching GloVe, and retains 91 TF-IDF features above its stated threshold.Source: 2) Comparing different word-embedding… · p. 8
  • GloVe performs best on RMSE and MAE, while TF-IDF outperforms GloVe on SMAPE; all three embeddings improve on the benchmark.Table IX presents forecasting errors for the benchmark and embedding-based models.Source: Table IX · p. 8, 2) Comparing different word-embedding… · p. 8

D. Application to NIE demand

  • Around 8% improvement over the benchmark is reported for all three metrics in NIE, compared with around 2% in the national case.The NIE benchmark uses demand lags, calendar features, and Belfast temperatures; Table X reports forecasting errors for textual feature groups added to that benchmark.Source: D. Application to NIE demand · pp. 8–9, Table X · p. 8
  • In NIE, WFT and SEB improve forecasting at both national and regional levels, while CFD, TDD, and GWED are also helpful.The feature groups were filtered again using the Granger test with NIE demand as the target.Source: D. Application to NIE demand · pp. 8–9
  • The NIE features relate load to Science&Environment and Wales news counts, daily sentiment dispersion, and topics spanning social events, politics, conflict, and culture.The paper describes the selected word- and topic-based features using their top words and latent dimensions.Source: D. Application to NIE demand · p. 9, Table XI · p. 9

IV. DISCUSSION

The discussion relates textual features to load through global and local analyses, with correlations varying by season, day, and hour, and local feature effects differing between days.

A. Global correlations

  • Correlations are more evident in spring, summer, and weekends; coronavirus-related words and Topic-18 correlate negatively with load most strongly in spring and summer, with weaker associations later.Driver-T is positively correlated with dawn and early-morning load in autumn and winter; Dim-51, representing UK train transportation, is positively correlated in spring and summer, especially during daytime and 19:00–22:00 in summer.Source: A. Global correlations · p. 9

B. Local correlations

  • LIME coefficients for coronavirus-T, coronavirus-D, and pandemic-D are negative on both selected days, while coefficients for other features vary between the days.Figure 7 compares the two days with the most negative and positive driver-T coefficients: 2021-02-20 and 2021-01-22.Source: B. Local correlations · p. 9

C. Causality effects

The paper examines time-varying causal effects of textual features on load at day-ahead and short-term horizons. Effects can differ by time of day, and only three features pass the short-term causality test.

1) Day-ahead forecasting:

  • “Pandemic” in news descriptions has a positive effect on demand at midnight but a negative effect during the rest of the day, illustrating time-varying effects that can reverse direction.The day-ahead analysis retains effects with p < 0.05 and τ > 0.10; Figure 8 plots selected feature coefficients across 48 half-hours.Source: 1) Day-ahead forecasting · p. 9, Figure 8 · p. 11
  • Topic-5 and Dim-51 have positive effects in the early morning and afternoon, sentiment negatively affects demand from morning to evening by around 15%, and Dim-9's effect exceeds 20% from noon to late night.Dim-51 represents transportation, while Dim-9 represents weapons.Source: 1) Day-ahead forecasting · pp. 9–10, Figure 8 · p. 11

2) Short-term forecasting:

  • Only power-B, pandemic-D, and Topic-5 pass the short-term causality test; power-B has a negative effect around 10%, while pandemic information and Northern Ireland politics can positively affect load.The short-term analysis retrains the causality model for midnight and 1 a.m.; pandemic information is positive in the second hour of the next day.Source: 2) Short-term forecasting · p. 10

D. Considerations

News-derived textual features improved UK electricity demand forecasts, and the paper examines how correlations and causality tests relate to those improvements. It also proposes explanations, notes study limitations, and identifies directions for further research.

D. Considerations

  • Dim-9 passed the Granger test despite low negative demand correlation yet showed relatively strong causality, while COVID-19 Topic-18 correlated more but could not cause a demand change.The examples show that Pearson correlation and causality effects do not align consistently, and Granger testing offers feature selection with limited causal explainability.Source: D. Considerations · p. 11
  • The study uses only BBC news and national aggregated demand, whose relatively stable pattern yields low day-ahead errors and makes textual-feature effects difficult to identify.The authors also note that the NLP methods are basic and suggest considering more advanced algorithms.Source: D. Considerations · p. 11

V. CONCLUSIONS

  • 4% in RMSE, 11% in MAE, and 10% in SMAPE were observed; the best model combined news-title word frequencies, sentiment scores, and GloVe embeddings from news bodies.Source: §V · p. 11
  • Table III compares improvements from adding textual features, while Figure 6 shows feature-load correlations across hours, seasons, and day types.Source: §V · p. 11
  • The forecasting chain converts news into numerical features, removes features with spurious correlations through a Granger test, then feeds retained features and known predictors into an existing model.Source: §V · p. 11
  • The authors link COVID-19 keywords to lockdown-era demand reductions, Northern Irish political news to demand reduction, and hypothesize oil-region tensions affect load through the economy.The international-tension mechanism is presented as an assumption; the COVID-19 account is framed as a possible relation to lockdowns in 2020–21.Source: §V · pp. 11–12
  • Future work could test longer horizons, finer spatial nodes, n-grams or deeper text models, social-media sources, and probabilistic forecasting.The authors also propose examining multilingual or geographically tagged social-media content and the distribution of error improvements across system nodes.Source: §V · p. 12
Loading 2301.07535v2…