Note: This repository is part of an ongoing project. At this stage, we are sharing selected portions of the codebase and a sample of the dataset. The complete source code and full private dataset will be made publicly available after the project's completion.

Model Architecture
The figure represents the end-to-end pipeline for the proposed VAE-S2S-BiLSTM-Encoder-Decoder model for short-term wind speed prediction. The architecture includes the following stages:

1. Data Preprocessing

• Inputs: Raw multivariate time series data consisting of temperature, humidity, wind speed, and power from Beijing’s weather.

Operations:

• Data cleaning: Removes noise and missing values.
• Normalization: Uses Min-Max scaling to rescale all variables to a 0 to 1 range.
• Temporal aggregation: Features like daily, weekly, and monthly patterns are extracted.
• Visualization: Heatmaps, line plots, and box plots are used to capture seasonal patterns and detect outliers.

2. Feature Engineering

• Purpose: Capture meaningful temporal dependencies and hidden correlations among variables.

Techniques:

• Seasonal and trend pattern analysis.
• Box plots and correlation heatmaps to explore dependencies between variables.
• Weekly and quarterly aggregations for modeling long-term trends.

3. Variational AutoEncoder (VAE) Module

• Encoder: Maps high-dimensional input features into a probabilistic latent space 𝑧 using neural networks.
• Latent Space Sampling: Uses mean and variance to generate latent variables via the reparameterization trick.
• Decoder: Reconstructs the original data from latent representations, learning abstract and nonlinear features.

🔹 The VAE addresses dimensionality reduction, overfitting, and nonlinearity.

4. Prediction Module – Deep Learning Models The latent features from VAE are passed into various prediction models:

✅ Standard LSTM

• Learns from sequential data, but processes information in a unidirectional manner.

✅ S2S LSTM Encoder-Decoder

• Learns to encode a sequence and decode it into another sequence. Suitable for multi-step forecasting.

✅ BiLSTM

• Processes the input sequence forward and backwards, capturing both past and future dependencies.

✅ S2S-BiLSTM Encoder-Decoder

• Integrates the strengths of bidirectional LSTM and sequence-to-sequence modeling.

• Supports multi-step and multi-feature time series prediction.

• Capable of capturing long- and short-term dependencies simultaneously.

🔹 This is the final architecture used in the Wind Speed Prediction App.

5. Output Layer

• Produces predicted values for wind speed.

• Evaluated using metrics like MAE, MSE, RMSE, MAPE, NMSE, and R².