Package 'WaveST'

Title: Wavelet-Based Spatial Time Series Models
Description: An integrated wavelet-based spatial time series modelling framework designed to enhance predictive accuracy under noisy and nonstationary conditions by jointly exploiting multi-resolution (wavelet) information and spatial dependence. The package implements WaveSARIMA() (Wavelet Based Spatial AutoRegressive Integrated Moving Average model using regression features with forecast::auto.arima()) and WaveSNN() (Wavelet Based Spatial Neural Network model using neuralnet with hyperparameter search). Both functions support spatial transformation via a user-supplied spatial matrix, lag feature construction, MODWT-based wavelet sub-series feature generation, time-ordered train/test splitting, and performance evaluation (Root Mean Square Error (RMSE), Mean Absolute Error (MAE), R-squared (R²), and Mean Absolute Percentage Error (MAPE)), returning fitted models and actual vs predicted values for train and test sets. The package has been developed using the algorithm of Paul et al. (2023) <doi:10.1007/s43538-025-00581-1>.
Authors: Dr. Md Yeasin [aut], Dr. Ranjit Kumar Paul [aut, cre], Akarsh Kumar Singh [aut]
Maintainer: Dr. Ranjit Kumar Paul <[email protected]>
License: GPL-3
Version: 0.1.0
Built: 2026-05-20 10:34:38 UTC
Source: https://github.com/cran/WaveST

Help Index

WaveSARIMA: Wavelet based Spatial ARIMA Model

Description

Builds wavelet sub-series features from lagged original data and spatially transformed data, then fits forecast::auto.arima() directly (no grid tuning). Uses wavelet features as regression variables (xreg) when they are available/valid.

Usage

WaveSARIMA(
  Data,
  Spatial_Matrix,
  Lag = 4,
  Wfilter = "d4",
  Wlevel = 4,
  Split = 0.8,
  y_cols = NULL,
  boundary = "periodic",
  fast = TRUE,
  seed = 123,
  verbose = TRUE
)

Arguments

Data

A numeric matrix or data.frame.
Spatial_Matrix

Square numeric matrix used for spatial transform (Bs).
Lag

Integer. Maximum lag used in tsutils::lagmatrix().
Wfilter

Character. Wavelet filter name passed to wavelets::modwt().
Wlevel

Integer. Wavelet decomposition level.
Split

Numeric in (0,1). Train ratio for time-order split.
y_cols

Integer vector. Target column indices in Data. Default = 1:k.
boundary

Character. Boundary for MODWT, default "periodic".
fast

Logical. Passed to MODWT; default TRUE.
seed

Integer. Random seed.
verbose

Logical. If TRUE, prints accuracy table.

Value

A list with:

  • model_table: selected ARIMA orders + AICc + number of regressors used

  • accuracy_table: train/test metrics per target

  • best_models: list of fitted auto.arima model objects per target

  • predictions: list of train/test data.frames (Actual vs Predicted) per target

  • error_log: errors if any target fails

References

Paul, R.K., Yeasin, M. and Manjunatha, B. (2025). WaveDeepCrop: wavelet enhanced deep learning framework for noise-resilient crop yield forecasting. Proceedings of the Indian National Science Academy, 91(4), 1310–1323.

Examples

Data <- matrix(rnorm(300), ncol = 3)
Bs <- matrix(c(0,1,0,
               0.5,0,0.5,
               0,1,0), nrow = 3, byrow = TRUE)
res <- WaveSARIMA(
  Data = Data,
  Spatial_Matrix = Bs,
  Lag = 4,
  Wfilter = "d4",
  Wlevel = 4,
  Split = 0.8,
  verbose = FALSE
)
res$accuracy_table

WaveSNN: Wavelet based Spatial ANN Model

Description

Builds wavelet sub-series features from lagged original data and spatially transformed data, then fits and hyper-tunes a feedforward neural network (via neuralnet) for one or multiple target columns.

Usage

WaveSNN(
  Data,
  Spatial_Matrix,
  Lag = 4,
  Wfilter = "d4",
  Wlevel = 4,
  Split = 0.8,
  hidden_grid = list(c(5), c(10), c(15), c(10, 5), c(15, 10)),
  threshold_grid = c(0.01, 0.001),
  stepmax_grid = c(1e+05, 2e+05),
  algorithm_grid = c("rprop+"),
  y_cols = NULL,
  boundary = "periodic",
  fast = TRUE,
  seed = 123,
  verbose = TRUE
)

Arguments

Data

A numeric matrix or data.frame. Non-numeric columns should be removed beforehand.
Spatial_Matrix

Square numeric matrix used for spatial transform (Bs). The function uses the first k columns of Data where k = min(ncol(Data), nrow(Spatial_Matrix), ncol(Spatial_Matrix)).
Lag

Integer. Maximum lag used in tsutils::lagmatrix().
Wfilter

Character. Wavelet filter name passed to wavelets::modwt() (e.g., "d4", "la8").
Wlevel

Integer. Wavelet decomposition level.
Split

Numeric in (0,1). Train ratio for time-order split.
hidden_grid

List of integer vectors. Hidden layer sizes to try, e.g. list(c(5), c(10), c(10,5)).
threshold_grid

Numeric vector. neuralnet convergence threshold values.
stepmax_grid

Numeric vector. neuralnet stepmax values.
algorithm_grid

Character vector. neuralnet algorithms, e.g. "rprop+".
y_cols

Integer vector. Target column indices in Data. Default = 1:k.
boundary

Character. Boundary for MODWT, default "periodic".
fast

Logical. Passed to MODWT; default TRUE.
seed

Integer. Random seed.
verbose

Logical. If TRUE, prints accuracy table.

Value

A list with:

  • hypertune_all: all hyperparameter runs for all targets

  • best_hyper_table: best hyperparameters per target

  • accuracy_table: best train/test metrics per target

  • best_models: list of fitted neuralnet objects per target

  • predictions: list of train/test data.frames (Actual vs Predicted) per target

References

Paul, R.K., Yeasin, M. and Manjunatha, B. (2025). WaveDeepCrop: wavelet enhanced deep learning framework for noise-resilient crop yield forecasting. Proceedings of the Indian National Science Academy, 91(4), 1310–1323.

Examples

Data <- matrix(rnorm(300), ncol = 3)
Bs <- matrix(c(0,1,0,
               0.5,0,0.5,
               0,1,0), nrow = 3, byrow = TRUE)

res <- WaveSNN(
  Data = Data,
  Spatial_Matrix = Bs,
  Lag = 4,
  Wfilter = "d4",
  Wlevel = 4,
  Split = 0.8,
  hidden_grid = list(c(5)),        # minimal to keep example fast
  threshold_grid = c(0.01),
  stepmax_grid = c(1e5),
  algorithm_grid = c("rprop+"),
  verbose = FALSE
)
res$accuracy_table