Abstract In this paper, we present FlareCast, a novel solar flare forecasting system developed under the framework of machine learning operations (MLOps) and made available as an online service to the community. FlareCast utilizes a deep Bayesian neural network that is specifically designed for forecasting solar flares. This neural network merges a convolutional neural network with long short-term memory to link the radial component of the vector magnetic field with anticipated maximum X-ray flux levels over the next 12–48 hr. Additionally, 24 hr forecasts are available for online access through our website. To address the uncertainties inherent in observational data, the model first predicts the probability distribution of maximal X-ray flux and then classifies these solar flares based on the predicted X-ray flux levels, rather than merely providing direct classification outcomes. Our trained Bayesian neural network achieves a precision of 97.9 for ≥X-class flares on the test set, outperforming existing operational models. To assess the model’s performance in practical applications, we adapt the MLOps concept. With the concept, we propose using the regression error of the maximal X-ray flux levels as a measurement criterion for the machine learning system. Additionally, we utilize attention maps from the prediction model, along with polarity reversal lines and magnetic field transition regions, to enable human users to monitor the system’s performance. FlareCast has been operational since 2024 January 1, consistently providing solar flare forecasting services.
Lv et al. (Wed,) studied this question.