ABSTRACT
This project presents a deep learning-based forecasting model for predicting weekly dengue cases using a combination of wavelet decomposition and long short-term memory (LSTM) networks. The objective is to enhance prediction accuracy by leveraging multi-resolution analysis and sequential learning techniques. Historical dengue case data from cities such as Ahmedabad, Iquitos, and San Juan is used to train and test the model. The data is first preprocessed by sorting it chronologically and applying smoothing techniques to reduce short-term noise. The time series is then decomposed using the Maximal Overlap Discrete Wavelet Transform (MODWT), which separates the original data into various frequency components. This decomposition helps isolate underlying patterns at different scales, making it easier for the model to learn both trend and seasonal behaviors. Each decomposed signal is normalized and converted into a supervised learning format. An individual LSTM model is trained on each component separately, allowing the network to specialize in learning specific patterns from different frequency bands. The predictions from all models are then summed to generate the final dengue forecast, ensuring that both long-term and short-term variations are captured effectively.
An ensemble of these LSTM models forecasts dengue cases for the upcoming weeks, and the results are post-processed to ensure values remain realistic. Trend direction is also determined to indicate whether dengue cases are expected to increase or decrease. The system further computes evaluation metrics such as MAE, RMSE, SMAPE, and MASE to assess model performance. The forecasting results are visualized through plots comparing actual and predicted cases. The model's ability to identify future trends and its relatively low error rates make it suitable for early warning systems and public health planning. This project is integrated with a web-based interface, allowing users to input forecast duration and receive real-time predictions


OBJECTIVES
 Develop a Hybrid Forecasting Model
To build a hybrid model integrating MODWT and LSTM to effectively forecast dengue cases. This will capture both non-linear temporal patterns and noise-filtered trends in the data.
 Enhance Prediction Accuracy
To improve forecasting accuracy by decomposing the original time series into meaningful components. MODWT helps reduce noise and enhances the learning capability of deep models like LSTM.
 Enable Trend Direction Classification
To classify the forecasted dengue trend as increasing or decreasing.
This helps public health authorities anticipate and respond to potential outbreaks in advance.
 Compare Model Performance with Metrics
To evaluate model performance using MAE, RMSE, SMAPE, and MASE.
These metrics help assess both absolute and relative accuracy for better validation.
 Implement Visualization and Interpretation
To provide intuitive visualizations for actual versus predicted cases.
This improves interpretability and helps stakeholders understand the disease trend.
 Support Public Health Decision-Making
To provide a practical forecasting tool deployable via a web interface.
This enables timely and informed intervention planning by health agencies.