Modeling Process

Convolutional Neural Network (CNN)

Convolutional Neural Networks (CNN) are a class of deep learning algorithms primarily used for processing data with a grid-like topology, such as images. In our case, we applied CNN to extract features from time-series data, attempting to detect spatial patterns that could relate to road conditions over time.

The CNN model performed well in feature extraction but faced challenges in capturing outliers and extreme cases in the data. It was particularly useful for identifying immediate road condition patterns but was less effective for forecasting future trends.

Recurrent Neural Network (RNN)

Recurrent Neural Networks (RNN) are designed to recognize patterns in sequences of data. Unlike CNNs, RNNs take the previous time step into account, which makes them ideal for tasks involving temporal data, such as weather patterns or road conditions.

The RNN model was able to capture the smaller details of the data, and it could also handle a bigger amount of it. However, it still had some issues on the lower bound of the data, predicting higher values than the actual ones. Overall accuracy is 81.32%.

Long Short-Term Memory (LSTM)

Long Short-Term Memory (LSTM) is a specialized type of RNN that is capable of learning and remembering over long sequences of data. This makes LSTM particularly effective for problems like road condition prediction, where dependencies span long periods of time.

The LSTM model showed very promising results, however we struggled with the feature engineering part, since the amount of hyperparameters to tune is very high. Overall a possible candidate for more serious future work.

Prophet

Prophet is an open-source tool developed by Facebook designed for time-series forecasting. It is particularly strong at handling seasonal data and missing values, which made it a valuable model for predicting road conditions based on weather trends and irregular patterns.

The Prophet model was a very quick and dirty attempt in the past couple of days. While we did not achieve very good results, our research showed that when properly tuned, Prophet can be a very powerful tool for time-series forecasting.