[Paper Review] Weather Analogs with a Machine Learning Similarity Metric for Renewable Resource Forecasting

Weiming Hu, Guido Cervone+2 more|arXiv (Cornell University)|Mar 7, 2021

Meteorological Phenomena and Simulations4 citations

TL;DR

This paper proposes a machine learning-based similarity metric for Weather Analog Ensemble (AnEn) forecasting that replaces traditional weighted Euclidean distance with a neural network-trained metric, enabling full use of all meteorological variables without prior feature selection or weight optimization. The method outperforms conventional AnEn in wind speed and solar irradiance forecasting by better correcting large errors and leveraging larger historical archives, with improved spatial prediction accuracy and robustness to model updates.

ABSTRACT

The Analog Ensemble (AnEn) technique has been shown effective on several weather problems. Unlike previous weather analogs that are sought within a large spatial domain and an extended temporal window, AnEn strictly confines space and time, and independently generates results at each grid point within a short time window. AnEn can find similar forecasts that lead to accurate and calibrated ensemble forecasts. The central core of the AnEn technique is a similarity metric that sorts historical forecasts with respect to a new target prediction. A commonly used metric is Euclidean distance. However, a significant difficulty using this metric is the definition of the weights for all the parameters. Generally, feature selection and extensive weight search are needed. This paper proposes a novel definition of weather analogs through a Machine Learning (ML) based similarity metric. The similarity metric uses neural networks that are trained and instantiated to search for weather analogs. This new metric allows incorporating all variables without requiring a prior feature selection and weight optimization. Experiments are presented on the application of this new metric to forecast wind speed and solar irradiance. Results show that the ML metric generally outperforms the original metric. The ML metric has a better capability to correct for larger errors and to take advantage of a larger search repository. Spatial predictions using a learned metric also show the ability to define effective latent features that are transferable to other locations.

Motivation & Objective

• To address the limitations of traditional AnEn similarity metrics that require heuristic feature selection and computationally expensive weight optimization.
• To improve forecast accuracy and calibration for renewable energy resources like wind speed and solar irradiance.
• To enhance robustness of analog-based forecasting under model updates and limited historical data.
• To enable effective spatial predictions by learning transferable latent features across grid points.
• To leverage modern ML architectures for more flexible and accurate weather analog identification.

Proposed method

• A triplet neural network architecture is trained to learn a similarity metric in a transformed latent space, using contrastive loss to group similar weather forecasts.
• The model is trained using a reverse analog technique that links model forecast errors to historical analogs, guiding the learning of predictive features.
• A pre-trained embedding network (LSTM-based) transforms raw NWP forecast variables into a lower-dimensional latent space where similarity is computed.
• The similarity metric is redefined in the latent space, allowing full use of all meteorological variables without manual weighting.
• The framework uses a large historical archive of NWP forecasts and observations to train the model, ensuring robustness to model changes.
• Spatial predictions are generated by applying the learned metric independently at each grid point, enabling scalable, high-resolution forecasting.

Experimental results

Research questions

• RQ1 Can a machine learning-based similarity metric outperform traditional weighted Euclidean distance in AnEn for renewable energy forecasting?
• RQ2 How does the ML metric perform in correcting large forecast errors compared to conventional AnEn?
• RQ3 To what extent can the learned metric generalize across spatial locations and improve gridded forecasts?
• RQ4 Does the ML-based approach reduce sensitivity to NWP model updates and improve tolerance to model parameter changes?
• RQ5 Can the latent features learned by the neural network capture physically meaningful weather patterns for improved analog selection?

Key findings

• The ML-based similarity metric significantly outperforms the conventional AnEn metric in forecasting both wind speed and solar irradiance.
• The ML metric demonstrates superior error correction capability, particularly for large forecast errors, due to its ability to leverage a broader set of meteorological variables.
• Spatial predictions using the learned metric show improved accuracy and better calibration compared to the optimal conventional AnEn.
• The method is more robust to NWP model updates and maintains performance over longer historical archives, unlike traditional AnEn which degrades under model changes.
• The latent features learned by the neural network are transferable across locations and effectively represent weather regimes relevant to renewable energy forecasting.
• Triplet networks with contrastive loss produce more coherent and physically meaningful clusters in the latent space than autoencoders, leading to higher-quality ensemble forecasts.