Multistep Input-Output Timeseries using LSTM

Probabilistic forecast a complex wind generator system with multiple parameters in assisting decision making and for predictive maintenance through data-driven techniques; in this case, recurrent neural networks.

Requirements

• python == 3.9.1
• tensorflow == 2.5.0
• pandas == 1.2.4
• numpy == 1.19.5
• seaborn == 0.11.1
• matplotlib == 3.4.2
• scikit-learn == 0.24.2
• keras == 2.4.3

Input Specifications

• Sample dataset is contained as data.csv
• inputs number of inputs for model training/testing
• Number of hidden neurons per LSTM and FC layers
• Moving Window of m number of time step inputs
• n day output Multistep forecasting

Script Specifications

• Load time series dataset CSV with specified (variables inputs inputs) – denoted in the sample dataset.
• Input preprocessed (StandardScalar) and using TimeSeriesSplit Cross-Validation
• Each LSTM model architecture has:

  • 2x LSTM layer (with their “number of hidden neurons” as variables) followed by 1x FC

  • Specifications:

    i. Moving Window Time Step Input i.e. where x multivariable inputs {x(t-m)…x(t-1) and x(t)}, where m is a variable

    ii. EarlyStopping Callback with patience = 20 during the training phase

    iii. Loss function: MSE

• Prediction phase to predict n days where y is output: y(t),y(t+1),…y(t+n)
• A multi-step approach, and where n is a variable

Implementation

Make changes in this part of the script to customise it to your dataset

obj = TSLSTM(
        path_to_csv="./data.csv", 
        inputs=[0,1,2,3,4,5], 
        neurons=50, 
        m=2, 
        n=1,
        approach="lstmd"
        )

To run the script, simply use python tslstm.py

Output

Two images will be stored in ./images and a logs.log file will be generated.

References

• Multistep Time Series Forecasting with LSTMs in Python
• Multi-Step Multivariate Time-Series Forecasting using LSTM

Contributions

I am looking forward to implementing the following 2 papers:

1. BAYESIAN RECURRENT NEURAL NETWORKS
2. Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning

Any contributions to the current script or in implementing the above 2 papers are welcome.