A New Approach to Using Neural Networks for Long-Term El Niño and La Niña Forecasting

This paper proposes a new approach to apply neural network and multiple linear regression (MLR) methods to predict the state of El Niño–Southern Oscillation (ENSO). The simulated parameters (Nino3, Nino3.4, and Nino4 indices) are divided into two components: low-frequency and high-frequency. Multilayer neural networks (MNNs) are used to simulate the high-frequency component, and MLR is used to study the low-frequency component. One feature of the approach is the detailed selection of predictors as input data to the model and subsequent simulating with search and verification of all possible constructions of the multilayer neural network. Based on the new approach, a model for forecasting extreme phases of ENSO phenomenon—Neural Network Model for ENSO forecast (NNM-ENSOv1)—is developed. The resulting model is characterized by low sensitivity to the spring predictability barrier. That is why the this model properties are significantly better compared to dynamic models. The forecast lead time of ENSO event state of this model is longer by 7 months. The model was verified over the control period from 2007 to 2022. NNM-ENSOv1 reproduces not only El Niño (EN) events, but also their type with a lead time of up to 1 year quite well. As a confirmation, we note that four out of five EN events, including their type, are predicted correctly. The probability of correctly identifying conditions typical of EN events is quite high and changes slightly within 76–83% when the forecast lead time changes within 11 months, while for La Niña (LN) the probability of correct identification decreases from 85 to 31% with an increase in lead time. Using the NNM-ENSOv1 model in November 2022, a forecast of the ENSO state in 2023 is constructed. This model successfully predicted the evolution of LN that occurred until February 2023, next, the subsequent neutral conditions in March–April and the onset of East Pacific EN from May 2023. As a result, the forecast for the first half of 2023 is confirmed. ENSO modeling results in real time are available on the model’s website at neuroclimate.com.