MufPrediction30min.jl

Julia application for real-time 30-minute-ahead Maximum Usable Frequency (MUF) prediction using pre-trained neural networks. MUF data is sourced from the EMBRACE ionosonde network and space-weather indices are computed via SpaceIndices.jl.

Overview

The application maintains a rolling 24-hour SQLite database populated with ionosonde MUF observations and space-weather indices (Dst, Kp, F10.7). A Flux.jl neural network, trained separately, takes these inputs and predicts the MUF 30 minutes into the future. Results can be displayed as an interactive prediction table or exported as a publication-quality PNG figure.

The workflow has four stages:

1. Setup — initialize the local database (once, idempotent).
2. Data — download recent MUF observations from EMBRACE and compute space-weather indices.
3. Prediction — run the neural network for a given station and timestamp.
4. Maintenance — prune database rows older than 24 hours.

Supported Stations

Station ID	Location	Latitude	Longitude
BVJ03	Boa Vista, RR	+2.87°	−60.71°
CAJ2M	Cachoeira Paulista, SP	−22.70°	−45.01°

Requirements

• Julia ≥ 1.11
• Dependencies listed in Project.toml (installed automatically by Julia's package manager)

Setup

Clone the repository, then run:

make install             # Instantiate the Julia environment
make initialize-database # Create the database (only needed once)

Usage

All entry points are available through make or directly from a Julia session.

Via Makefile

make update-all        # Download MUF data + compute space indices for all stations
make predict           # Run the neural network — interactive station menu
make plot              # Generate and save the prediction figure — interactive station menu
make plot-day          # Generate and save the daily observed vs. predicted figure — interactive menus
make cleanup-database  # Remove rows older than 24 hours

Run make (or make help) to list all available targets.

Via Julia REPL

using MufPrediction

# First-time setup.
initialize_database()

# Populate the database with recent data.
update_all()

# Predict the MUF 30 minutes ahead (interactive station menu).
predict_muf_30min()

# Or supply a specific model and timestamp.
predict_muf_30min("models/nn-BVJ03-2026-06-15T13:17:55.342.jld2", DateTime("2026-06-15T12:00:00"))

# Plot the 12-hour history and 30-minute forecast (interactive station menu).
plot_and_save_muf_prediction()

# Plot the observed vs. predicted MUF for one full UTC day (interactive menus, or pass the
# model and the date directly).
plot_and_save_muf_day()
plot_and_save_muf_day("models/nn-BVJ03-2026-06-15T13:17:55.342.jld2", Date(2026, 8, 11))

# Periodic maintenance.
cleanup_database()

Neural Network Models

Pre-trained model archives (.jld2) are stored in the models/ directory. Each archive encodes the station identifier in its filename:

models/nn-BVJ03-<timestamp>.jld2
models/nn-CAJ2M-<timestamp>.jld2

The station is read directly from the archive, so no station argument is needed at inference time. Models are trained by the companion train_neural_network_for_single_station package.

Input Features

The network receives a 14-dimensional feature vector:

Feature	Description
doy_sin, doy_cos	Day-of-year (cyclically encoded)
hour_sin, hour_cos	Hour of day (cyclically encoded)
cos_sza	Cosine of the solar zenith angle
Dst, Dst_lag_60min	Geomagnetic Dst index (current and 60 min prior)
Kp	Geomagnetic Kp index
F10obs, F10avg	Observed F10.7 and 81-day average F10.7 solar flux
MUF	Current MUF observation
MUF_trend_20min, MUF_trend_40min, MUF_trend_60min	MUF trends over the past 20, 40, and 60 minutes

Output

plot_and_save_muf_prediction saves a PNG to the current directory with the filename:

MUF_+30min_prediction-<Station_Name>-<yyyy-mm-dd_HHMM>UTC.png

The figure shows the observed MUF (solid line), the rolling 30-minute-ahead predictions over the past 12 hours (dashed line), and the operational forecast point for the next 30 minutes (star marker).

plot_and_save_muf_day saves a PNG with the filename:

MUF_daily_+30min-<Station_Name>-<yyyy-mm-dd>.png

The daily figure covers one full UTC day for historical verification: the observed MUF (solid line) and the rolling 30-minute-ahead predictions (dashed line), plus the MAE / RMSE of the predictions in the information panel. Since the day is in the past, no operational forecast is drawn. The figure is styled with the SatelliteAnalysis.jl Makie theme (light by default; pass theme = :dark for the dark variant).

Data Source

MUF observations are downloaded from the EMBRACE ionosonde data archive operated by INPE:

https://embracedata.inpe.br/ionosonde/{STATION}/

The database follows a "DB primary, EMBRACE fills gaps" strategy: only files absent from the local database are downloaded, making frequent calls fast once the database is warm.