Deep-learning-based prediction of Precipitable Water Vapor in the Chajnantor area

TL;DR

This work introduces a site-specific LSTM model to forecast precipitable water vapor (PWV) at the Chajnantor plateau using two 183 GHz radiometers and local meteorological data. The model predicts PWV for $12$, $24$, $36$, and $48$ hours ahead, showing substantial improvements over the Global Forecast System (GFS) for the short horizons ($12$–$24$ h) with RMSE reductions of about 50% and MAPE around $22$%. Performance declines at longer horizons due to error accumulation and unresolved small-scale atmospheric processes, highlighting opportunities for data augmentation, additional inputs, and alternative architectures. The study demonstrates the practical potential of deep learning for rapid, site-specific PWV forecasting to optimize observational scheduling at high-altitude radio observatories and provides a pathway for extending the approach to other sites.

Abstract

Astronomical observations at millimeter and submillimeter wavelengths heavily depend on the amount of Precipitable Water Vapor (PWV) in the atmosphere, directly affecting the sky transparency and degrading the quality of the signals received by radio telescopes. Predictions of PWV at different forecasting horizons is crucial to support telescope operations, engineering planning, and observational scheduling and efficiency of radio observatories installed in the Chajnantor area in northern Chile. We developed and validated a Long Short-Term Memory (LSTM) deep learning-based model to predict PWV at forecasting horizons of 12, 24, 36, and 48 hours using historical data from two 183 GHz radiometers and a weather station in the Chajnantor area. We find the LSTM method is able to predict PWV in the 12 and 24 hours forecasting horizons with Mean Absolute Percentage Error (MAPE) of 22% compared to 36% of the traditional Global Forecast System (GFS) method used by Atacama Pathfinder EXperiment (APEX) and the Root Mean Square Error (RMSE) in mm are reduced by 50%. We present a first application of deep learning techniques for preliminary predictions of PWV in the Chajnantor area. The prediction performance shows significant improvements to traditional methods in 12 and 24 hours time windows. We also propose upgrades to improve our method in short (< 1 hour) and long (> 36 hours) forecasting timescales for future work.

Figures (9)

• Figure 1: Correlation matrix for PWV APEX, PWV UCSC, RH, Temp, WS, and $U$- and $V$-wind components. Warmer colors are positive correlations, while cooler colors are negative correlations.
• Figure 2: Number of occurrences for each variable after data cleaning, imputation of missing values, and 3-hour averaging.
• Figure 3: Power spectral density plots from the FFT analysis of PWV APEX, PWV UCSC, RH, and Temp. The x-axis represents frequency in cycles per year.
• Figure 4: Diagram of the LSTM architecture.
• Figure 5: Comparison of PWV APEX (blue) and PWV LSTM (orange) for the different forecasting horizons.