Project - Sunspot Prediction Using Machine Learning

Explore sequence models for time series data, including recurrent neural networks and backpropagation through time, and examine vanishing gradients that limit long-term memory in many-to-many and many-to-one setups.

Understand vanishing gradients in RNNs and how long short-term memory networks use forget, input, and output gates to retain information in the cell state; compare with gated recurrent units.

Switch the Kaggle notebook accelerator from none to gpu to speed up model training, while watching the 33-hour weekly quota and rerun the notebook after enabling gpu to avoid errors.

Manipulate a one-dimensional time series into a three-dimensional input for lstm using the window method, with a sunspot data example to forecast the next value.

Understand how to manipulate a sunspot time series with a window dataset, creating sliding input windows and next-value labels for LSTMs, using batch and prefetch.

Split the sunspots time-series data into training and validation sets with a 90/10 train-test split to build an LSTM forecast predicting the next value from a three-value window.

This lecture outlines key LSTM parameters for sunspot prediction, including window size, batch size, shuffle buffer size, training epochs, and momentum SGD, with univariate time-series data.

Understand Huber loss as a robust loss for sunspot prediction, blending mean squared error and absolute error via a delta threshold, with Keras implementation guidance.

Build a hybrid conv1d-lstm model with dense and lambda layers for sunspot prediction, and use a learning-rate scheduler with SGD and momentum to identify the optimal rate.

Identify the optimal learning rate from training history and minimum loss, then train the model with that rate, tracking losses per epoch and plotting to evaluate performance.

Decode the model forecaster function and apply it to test data. Assess predictions against validation data using mean absolute error.

Explore solar data, visualize cycles, and build a sequence model using lstm or gru, tuning parameters to reduce test data error and guide future Astronomical Data Science projects.