CO₂ Emissions Forecasting with Shallow Neural Nets in Python

Prepare time-series data from the World Bank CO2 dataset, cleaning NaN values. Convert year to numeric, set a time index, pivot by country, and export to Excel for modeling.

Define the target variables and split the 44-row dataset into 35 training samples and 9 test samples, using 3 feature columns across eight regional targets indexed by year.

Scale the data to improve convergence, reduce overfitting risk, and boost numerical stability for shallow and deep neural networks. The scaling step is not required for linear regression or ARIMA.

Create and compile region-specific shallow neural network models in TensorFlow, with a single 100-neuron hidden layer, relu activation, mean squared error loss, and the Adam optimizer.

Fit eight shallow neural network models for time series forecasting by configuring three layers and training with 100 epochs and batch size 8 using Adam to minimize mean squared error.

Learn to draw a shallow neural network model and troubleshoot the plot_model function with pydot and Graphviz, including installation steps, path issues, and restarting Jupyter for successful execution.

Explain activation functions in a shallow neural network for time series forecasting, highlighting a nonlinear relu hidden layer and a linear output layer to capture nonlinear patterns.

Generate predictions for time series with shallow neural networks by training on scaled inputs and unscaling outputs via inverse_transform. Produce training and test set predictions for each region.

Examine overfitting in shallow neural networks by comparing test and training MAPE, using test minus training error to flag regions like China and South Africa against a 10% benchmark.

Assess overfitting first, then compare the SNN's test MAPE with a naive one-period-ahead model using World Bank data; accept forecasts only if the SNN outperforms the naive model.

Explain hyperparameter tuning and sensitivity analysis for time series forecasting, using training, validation, and test sets to assess overfitting and test MAPE forecasting error.

Conduct sensitivity analysis on shallow neural networks by varying polynomial degree, layers, neurons, activation, batch size, learning rate, and epochs, and measure test set MAPE as forecasting error.

Distinguish forecasts from predictions, apply polynomial expansion and scaling to World Bank CO2 data (1971–2014), evaluate with training and test errors, and prepare 2015–2050 forecasts.