Introduction
The AFATM TSAR module is Predico’s full machine learning forecasting utility. It will provide both univariate and multi-variate forecasting capability
Non Seasonal Models
We have used non-seasonal models in reservoir forecasting because they offer greater flexibility and accuracy in capturing the underlying trends and dynamics of reservoir production. Unlike seasonal models, which assume periodic patterns and fixed seasonal variations, non-seasonal models do not rely on predefined periodicity or specific seasonality assumptions. This flexibility allows them to adapt to complex and evolving production behaviours, such as sudden changes in reservoir conditions, operational adjustments, or varying external factors.
NBEATS (Neural Basis Expansion Analysis for Time Series)
NBEATS is a deep learning model designed to forecast directly from time series data without relying on traditional decomposition methods. It features a stack of blocks, each consisting of fully connected layers, and these blocks are divided into two types: trend and seasonality blocks. Each block processes the input time series to generate both trend and seasonal forecasts. The model combines the outputs from multiple blocks to produce a comprehensive final forecast. This structure enables NBEATS to learn complex patterns and interactions across different time horizons without explicit seasonality or trend specifications, making it adaptable to a variety of data types.
NBEATS-X (Extended NBEATS)
NBEATS-X builds on the original NBEATS architecture by introducing enhancements to improve forecasting accuracy and flexibility. It incorporates additional features, such as external covariates (e.g., weather conditions or economic indicators), which provide extra context for the forecasting process. The extended architecture allows the model to integrate these additional features, enhancing its ability to capture relationships between the primary time series and external influences. This makes NBEATS-X particularly adept at handling complex datasets with external variables that impact forecasting outcomes.
NHITS (Neural Network-based Hierarchical Forecasting)
NHITS extends the NBEATS model by incorporating hierarchical forecasting techniques to handle data at multiple levels of granularity. It processes time series data through a hierarchical structure, which allows it to capture both fine-grained and broad patterns. The model generates forecasts at different scales and aggregates them to produce a comprehensive prediction. This multi-scale learning approach enables NHITS to effectively capture both short-term fluctuations and long-term trends, providing a robust framework for analyzing complex time series data.
MLP (Multi-Layer Perceptron)
MLP is a fundamental neural network model consisting of multiple layers of interconnected neurons. Each layer transforms the input data through weighted sums and non-linear activation functions. The architecture includes an input layer, one or more hidden layers, and an output layer, with the hidden layers learning complex patterns through back propagation. MLPs adjust their weights based on the error between predicted and actual values, refining their predictions iteratively. This process enables MLPs to model intricate relationships within the data, making them a versatile tool for various forecasting tasks.
KAN (Kolmogorov–Arnold Networks)
KAN is a specialized model based on chaos theory, designed to handle highly complex and chaotic time series data. By leveraging principles from chaos theory, KAN captures the dynamic and unpredictable nature of systems with irregular behaviour. It models the inherent complexity of time series data by understanding its chaotic structure and uses dynamic learning mechanisms to predict future values. This approach allows KAN to effectively manage data that traditional models might struggle to handle, making it suitable for forecasting in highly volatile and chaotic environments.