Multi-Resolution, Multi-Horizon Distributed Solar PV Power Forecasting with Forecast Combinations

This repository includes the code for the paper titled as "Multi-Resolution, Multi-Horizon Distributed Solar PV Power Forecasting with Forecast Combinations". Please cite the following paper, if you are using this repository.

@article{Perera2022,
    title = {{Multi-Resolution, Multi-Horizon Distributed Solar PV Power Forecasting with Forecast Combinations}},
    year = {2022},
    journal = {Expert Systems with Applications},
    author = {Perera, M.,de Hoog, J., Bandara, K.,Halgamuge, S.},
    url = {https://doi.org/10.1016/j.eswa.2022.117690},
    doi = {10.1016/j.eswa.2022.117690}
}

Software Requirements

Software requirements are listed in requirements.txt

Execution Instructions

A example of running the code is in main.py for small sample dataset provided in the data folder.

Running the code

python main.py

The following is an example output of the main.py. It will provide the mean MASE across all test samples and a visualisation of the forecasts produced by all approaches.

Mean Absolute Scaled Error (MASE) for the test samples

{
    "sn": 0.4617352375226971, 
    "(s)arima": 0.8710616362438794,
    "(s)arimax": 0.7290526167551878, 
    "mlr": 0.9119233826493682, 
    "svr": 0.8491110380174636, 
    "pso- unconstrained": 1.0867953618937825, 
    "pso [0,1]": 0.4541319252235865, 
    "pso- convex": 0.6832882493648744, 
    "average": 0.6411990407241464, 
    "re": 0.7161379879466082
}

Visulatisation of the forecasts produced by all approaches

Directory Structure

|-main.py - example code to run all forecast approaches and visualise
|-data - data folder
|-constants.py - constants related to the project
|-plot.py - functions related to plotting
|-run_base_models.py - code to run all base learners
|-run_combinations.py - code to run all forecast combination methods
|-timeseries_split.py - class to split time series data
|-forecast_combinations_approach - code for the forecast combination approach (this runs the run_bse_models and run_combinations)
|-util.py - additional functions
|-combinations - class implemenations for average, pso methods and recursive ensemble
|-machine_learning - class implementations for linear and support vector regression models
|-naive_models - class implementation of the seasonal naive model
|-tsmodels - class implenation of the (s)ARIMA and (s)ARIMAX models

Executing With Your Own Data

1. Add your data to the data folder as per the provided example
2. Add the corresponding external variables to the data folder as per the provided example
3. Update the HORIZON_INFO dictionary in constant.py. This dictionary contains the following parameters.

   '1D' - A string identifier corresponding to the forecast horizon. Has to be a python recognized timedelta format
   'resolution' - A string identifier for the resolution of the data (e.g. H, min, D)
   'horizon_as_int' - An integer value corresponding to the number of steps to forecast in the horizon
   'resolution_as_str' - A string identifier corresponding to the resolution of the data. Has to be a python recognized timedelta format
   'seasonality' - An integer value for the seasonal frequency of the data
   'arima_params': {
           'seasonal_freq': An integer value noting the seasonal frequency of the data for a SARIMA model
           'seasonality': A boolean value indicating whether to fit a non-seasonal or seasonal model (True if a seasonal model is required, else False)
           'fourier': A boolean value whether to fit a fourier series as the seasonality (important when dealing with long seasonalities)
           'fourier_terms': None or an integer indicating the number fourier terms when the seasonality is modelled as a fourier series
           'maxiter': None or an integer indicating the number of iterations to perform when updating an ARIMA/SARIMA model
       }

Notes:

The values for the following parameters have been reduced for a faster execution. Increase the value of these parameters to a value > 100 when executing the code with large amount of data.

Class pso_model.py - iterations, hyper_parameter_iter

Change the number of training, testing days passed to the TimeSeriesDivide function based on your desired evaluation period