PaperlayerView exact revision on arXiv

Source-linked AI summary

Multistep and Multivariate Time Series Explainability with GradientSHAP: A Wind Power Case Study

Robbie Southam, Zhong Fan, Ke Li

arXiv:2610.40330v1cs.AIcs.LG

TL;DR

Short-term wind forecasting matters for grid operations, but opaque AI models can hinder trust and adoption. This paper applies GradientSHAP with tailored visualizations to multivariate, multistep wind forecasting, reporting temporal and site-specific patterns in model explanations.

  • Problem

    Opaque forecasting models can hinder stakeholder trust and AI adoption, while explainability research on multivariate, multistep time series remains limited.

    Source: Highlights · p. 2, §1 · p. 4

  • Method

    The paper adapts GradientSHAP for multivariate, multistep wind forecasting and uses visualizations to interpret feature contributions across lags and prediction steps.

    Source: Highlights · p. 2, §4.1 · p. 9

  • Results

    Explanations show that recent power is most predictive for the first seven forecast steps, after which site ID and site-specific factors become more influential in mid-range steps.

    Source: §4.1 · p. 9

  • Takeaways & Limitations

    The authors report that GradientSHAP combined with domain knowledge can provide insights into model behaviour and support more informed decision-making.

    Source: Highlights · p. 2

    Interpretation depends on substantial domain expertise, while static visualizations offer limited actionable dynamic insight and sampling choices may introduce bias.

    Source: §5 · p. 17

Abstract

from arXiv · show

As the effects of climate change become increasingly apparent, the demand for renewable energy sources such as wind power continues to grow. However, renewables' variable and intermittent nature presents significant challenges in their integration into energy grids. Artificial intelligence (AI) has emerged as a powerful tool to address these issues. Still, its adoption in the energy sector has been hindered by the opacity of many AI models, which often function as black boxes. This paper introduces a carefully adapted GradientSHAP approach to provide interpretability for multivariate, multistep wind power forecasting -- a largely underexplored area. This capability is critical for day-ahead grid management and, to the best of our knowledge, this is the first study to apply GradientSHAP in the context of wind power forecasting. To address the inherent dimensionality, we developed novel visualizations that deliver both local and global explanations, uncovering intricate relationships between environmental factors, site-specific conditions, and temporal dynamics. While GradientSHAP is not an out-of-the-box solution, our findings demonstrate that, with thoughtful handling of SHAP values and domain expertise, it can offer valuable insights into model behaviour, supporting more informed and trustworthy decision-making. Nevertheless, its complexity poses challenges for practical implementation. Future research should focus on streamlining these interpretability methods to make them more actionable for routine deployment in critical energy applications.

Highlights

The paper applies GradientSHAP to day-ahead wind power forecasting and develops visualizations and guidance for interpreting complex time-series explanations. The analysis reports temporal, spatial, and feature-specific patterns in model behaviour.

Highlights

  • The study applies GradientSHAP to day-ahead wind power forecasting, addressing a gap in explainable AI for renewable energy systems.Source: Highlights · p. 2
  • A practical guide combines domain knowledge with visualizations tailored to complex time-series data to derive insights from SHAP methods.Source: Highlights · p. 2
  • GradientSHAP analysis reveals temporal, spatial, and feature-specific dynamics, validating known relationships and uncovering novel interactions in wind power forecasting.The paper presents this as a novel application of GradientSHAP to day-ahead forecasting.Source: Highlights · p. 2
  • The work advances research on multivariate, multitarget time-series explainability, which the authors describe as important for renewable forecasting and operational transparency.Source: Highlights · p. 2

A Wind Power Case Study

This front matter identifies the paper’s authors, their University of Exeter affiliations, and the paper’s subject keywords.

A Wind Power Case Study

1. Introduction

The introduction motivates explainability for short-term wind forecasting and positions GradientSHAP as an approach for a relatively unexplored multivariate, multistep setting. It previews explanations aligned with real-world dynamics while noting the need for expert knowledge.

1. Introduction

  • Wind-power variability challenges grid integration, while short-term forecasting supports load management, reliable dispatch, and renewable penetration.Source: §1 · p. 4
  • Opaque forecasting models can hinder stakeholder trust and AI adoption in high-stakes energy settings.Source: §1 · p. 4
  • The paper presents what it calls the first GradientSHAP application to wind forecasting and extends work on multivariate, multistep time-series explainability.It reports that model predictions rely strongly on diurnal patterns and, to a lesser extent, seasonal trends.Source: §1 · p. 4
  • GradientSHAP’s reliance on expert knowledge limits its broader applicability, despite its interpretability potential when combined with domain expertise.Source: §1 · p. 4
  • The authors report that their analysis largely exposes model decisions and finds significant alignment with real-world wind-forecasting dynamics.They connect transparency to more informed decisions by grid operators and policymakers, as well as grid stability and operational efficiency.Source: §1 · p. 4

2. Related Work

Prior explainability work in energy forecasting has concentrated more on demand-side tasks, with relatively few studies addressing multivariate and multistep time series together. The paper positions its GradientSHAP application as a response to that gap.

2. Related Work

  • Most SHAP energy-forecasting studies focus on demand-side tasks; few address multivariate or multistep series, and even fewer address both.Source: §2 · p. 5
  • A load-forecasting study found SHAP useful for identifying features that degraded accuracy, but reported inaccurate importance rankings likely tied to DeepSHAP assumptions.Source: §2 · p. 5
  • A heat-demand study used DeepSHAP for multivariate, multistep explanations, but noted domain-expertise needs and limits from engineered features and DeepSHAP’s independence assumption.Source: §2 · p. 5
  • Energy-domain GradientSHAP studies are scarce, and omitted background-sample counts complicate method comparisons.Source: §2 · p. 5
  • The paper responds by applying GradientSHAP to multivariate, multistep energy forecasting and aims to provide an example for domain experts.Source: §2 · p. 6

3. Methodology

The methodology pairs a multistep wind-power forecasting setup with an adapted GradientSHAP analysis, from data preparation and attention-based Seq2Seq modeling to feature attribution.

3.1. Dataset

  • The WIND dataset contains seven years of five-minute meteorological data from more than 126,000 sites.Source: §3.1 · p. 6

3.1.1. Feature Selection

  • Four states supplied 12 selected sites; the hourly data used 48 historical lags, and hourly and monthly trends were most prevalent.Data were split temporally using 2008–2010 for training, 2011 for validation, and 2012 for testing; correlation and stationarity tests informed lag selection and interpretation.Source: §3.1.1 · p. 6

3.1.2. Preprocessing

  • Cyclical encoding was applied to day, month, and wind direction, while rolling windows paired 48 historical time steps with 24 prediction steps.Site IDs were integer encoded.Source: §3.1.2 · p. 6

3.2. Model Architectures

  • The forecasting architecture uses an LSTM encoder-decoder with Luong attention to weight input time steps when generating each prediction.The paper compares dot, general, and concat alignment scores; attention focuses on the recent 48-hour context.Source: §3.2 · p. 7

3.3. Optimisation

  • The general attention scoring function performed best, and the encoder-decoder architecture generally performed best with attention.Bayesian Optimisation was used to select architecture and remaining parameters; early stopping and a learning-rate scheduler were also used.Source: §3.3 · p. 7

3.4. Explainability by Feature Attribution

  • GradientSHAP supports explanations across lags, features, prediction steps, and their interactions in this multistep, multivariate forecasting setting.It combines expected gradients with sampling across multiple baselines and averaging gradients along paths between each baseline and the input.Source: §3.4 · pp. 8–9

3.4.1. Utilising GradientSHAP

  • GradientSHAP explanations were generated for 300 samples using a background set of 5,000, with hooks restoring gradient flow for the integer-encoded site embedding.The explanations cover each feature across 48 historical time steps and 24 future predictions.Source: §3.4.1 · p. 9

3.5. Experimental Setup

Training, inference, and explanation generation used a specified GPU, CPU, and memory configuration.

3.5. Experimental Setup

  • Training, inference, and explanation generation ran on an Nvidia GeForce GTX 1080 Ti, Intel Core i7-14700k, and 48GB RAM.PyTorch was used for model development, training, and GPU inference.Source: §3.5 · p. 9

4. Results & Discussion

GradientSHAP visualizations show that wind-power forecasts rely on recent power, site-specific conditions, and changing temporal patterns. The analyses validate expected relationships while identifying site-level variation and feature interactions that qualify those patterns.

4.1. Summary Plots

  • Power at t-0 drives short-range predictions, while site ID becomes more influential in mid-range horizons; temperature is negligible and earlier lags generally contribute little.The power attribution is predictive mainly for the first seven prediction steps.Source: §4.1 · p. 9, Figure 1 · p. 10
  • For one peak-production instance, site ID becomes more important as power becomes less important; positive site ID and hour attributions indicate higher predicted output.The authors interpret stable periods as relying more on power, wind speed, and monthly seasonality, and unstable periods more on diurnal and site-specific factors.Source: §4.1 · p. 9, Figure 2 · p. 11
  • The layered violin plot ranks the top 10 features: power at t-0 is most important, with high values supporting positive predictions and low values negative ones.Hour at t-0 and t-1, wind speed at t-0, and wind direction at t-0 are also significant.Source: §4.1 · pp. 9–10, Figure 3 · p. 11

4.2. Analysing Interactions

  • Site-level SHAP values vary: Washington and Texas are higher, while New York and Iowa sites are lower, suggesting weaker variability or environmental interactions there.High wind speeds are concentrated at sites 2, 9, and 10 but are not directly influencing the model; wind direction may either enhance or reduce site ID’s importance depending on local patterns and turbine alignment or yaw control [46].Source: §4.2 · pp. 10–11, Figure 4 · p. 12

4.3. Verifying Relationships

  • Joint plots validate a positive wind-speed/SHAP relationship but reveal clusters where high wind speeds coincide with low SHAP values, potentially reflecting suboptimal wind direction or other confounders.The plots pair primary and secondary features and use density regions; the authors also suggest temperature-related stability may reduce wind-speed and power attribution, while noting pressure data and cluster-to-site identification could improve interpretation.Source: §4.3 · pp. 11, 14

4.4. Temporal Analysis

  • Temporal analyses show diurnal and seasonal shifts: site ID peaks in early hours, power and wind speed peak at selected hours, and both dip seasonally from March to June.Temperature dominates some earlier hours and peaks in winter, while power and wind speed peak in October; the authors caution that hourly patterns are sporadic and normalization may minimize per-feature trends.Source: §4.4 · pp. 14–16

4.5. Summary

  • The paper’s synthesis finds site ID important across lags and horizons, with t-0 dominant for most variables and site ID increasingly influential at mid-range horizons.The summary also highlights non-linear interactions, temperature-related variability, and cases where wind direction may override wind speed under suboptimal alignment or local conditions.Source: §4.5 · pp. 16–17
  • Overall, global and local visualizations validate known relationships, reveal temporal patterns, and identify interactions that inform model predictions.The authors describe the analysis as aligning model decision-making largely with real-world dynamics.Source: §4.5 · p. 17

5. Conclusion

GradientSHAP analysis of day-ahead wind power forecasting revealed global, temporal, and interaction patterns, while its interpretation remains context-dependent and requires substantial domain expertise.

5. Conclusion

  • Site ID was the most important global feature, while power at t-0 dominated short-term predictions.Site ID remained important across time steps, particularly at mid-range horizons; temporal heatmaps also showed diurnal and seasonal patterns.Source: §5 · p. 17
  • The analysis highlighted interactions involving site ID, wind direction, and temperature, and validated known relationships.The authors relate the patterns to atmospheric stability and pressure gradients and say validation strengthened confidence in robustness and alignment with real-world dynamics.Source: §5 · p. 17
  • GradientSHAP’s interpretability is context-dependent and requires substantial domain expertise to translate SHAP values into insights.Static visualisations provided explanatory power but limited actionable dynamic insights; out-of-manifold samples and fixed sampling methods may have introduced biases.Source: §5 · p. 17
  • Future work could add atmospheric pressure or spatial data and develop interactive explanation tools and domain-informed feature clusters.The paper also proposes comparing SHAP values with model attention weights.Source: §5 · p. 17

Appendix A.1. Supplementary Results

The appendix includes a temporal heatmap of one sample’s features across lags.

Appendix A.1. Supplementary Results

  • Figure A.7 presents a temporal heatmap of one sample’s features across lags.Source: Figure A.7 · p. 18

Appendix A.2. Model And Optimisation

The appendix lists summaries of the final model and its hyperparameters, fixed hyperparameters, and model architecture.

Appendix A.2. Model And Optimisation

  • Table A.1 summarizes the final optimized model and training hyperparameters.Source: Table A.1 · p. 19
  • The model has 46,090 trainable parameters and 1.43 million total mult-adds.The reported model has no non-trainable parameters.Source: Appendix A.2 · p. 19
  • Table A.2 summarizes the final forecasting model architecture, and Table A.3 lists fixed hyperparameters.LR denotes learning rate in Table A.3.Source: Table A.2 · p. 19, Table A.3 · p. 20

Appendix A.3. Preprocessing

The appendix illustrates three non-overlapping windows for site 0.

Appendix A.3. Preprocessing

Related papers

Loading 2610.40330v1…