GNGTS 2022 - Atti del 40° Convegno Nazionale

GNGTS 2022 Sessione 1.2 119 FORECASTING SOIL CO2 FLUX AT MT. ETNA VOLCANO S. Scudero, A. D’Alessandro Istituto Nazionale di Geofisica e Vulcanologia, Osservatorio Nazionale Terremoti, Rome, Italy Introduction. Procedures for forecasting time series are various and the choice of the most suitable technique depends mainly on the type of data. In this contribution we deal with a very peculiar type of data. The monitoring of CO 2 flux may provide information about the magmatic processes and represent useful eruption predictors (Aiuppa et al. , 2007). For this reason, also the monitoring of diffuse soil CO 2 has become very important in recent years. At Mt. Etna Volcano (Sicily, Italy) a network of 14 sites records hourly measures of soil CO 2 (Gurrieri et al. , 2008; Liuzzo et al. , 2013); the data coming from all the sites, after a processing, are summarized to provide a unique, combined signal which is used as one of the indicators of the state of activity of the surveillance of the volcano. The procedure leading to the unique CO 2 signal consists in four steps: i) reduction from hourly to daily time series; ii) filtering of the periodic components at frequency > 7 day -1 after FFT analysis; iii) multiple linear regression and FFT filtering the atmospheric influences (between 300 and 400 day -1 ); iv) normalization of each CO 2 flux series in the range 0-1 as preparatory step to obtain a unified signal of CO 2 flux from all the monitoring stations. Such procedure was, proposed by Liuzzo et al. , (2013), then further developed and validated during several eruptive episodes (Cannata et al., 2015; Gurrieri et al. , 2021; Paonita et al. , 2021; Scudero et al. , 2022). In this paper we tested the short-term performance forecast of this unified CO 2 signal. Method. Aiming at forecasting this unified signal we should keep in mind how it was previously processed and filtered. It can be considered as a univariate time series, and differently from the methods for multivariate series which are based on correlation, the univariate forecast relies on time-dependence. A time series can be considered as the combination of an auto-correlated part related to the measure of the physical phenomenon, and stochastic/random part. It can be decomposed as the sum of a mean value and an error which cannot be explained by the mean. The former has been estimated by means of a moving window; in particular, we adopted a linearly weighted moving average in which the weight of each value within the window decreases in arithmetical progression moving away from the value to predict. The latter value, considering that data have been already filtered to remove the seasonal components, can be well approximated by a withe noise (WN). WN is characterized by equal intensity across all allowable frequencies; it has been simulated with an ARIMA( p,d,q ) model in which all the three components (i.e. the autoregressive order p , the order of integration d , and the moving average order q ) are equal to zero. The standard deviation of the WN over a given window width provide the stochastic component of the signal. The appropriate values for the width of the windows (either for the mean and for the stochastic component) have been chosen after performing some tests. In particular, we performed some forecast simulations on portions of known signals and compared the density estimates of each time series before and after the forecasts. The comparison has been performed by means of a bootstrap hypothesis test of equality of the pairs of density estimations. The selected windows’ widths, which minimize the differences between the densities, result proportional to the chosen steps of the forecast, namely 1, 5 and 10 days (Tab. 1). As the error estimation, the root mean square error (RMSE) and the determination coefficient (R 2 ) have been calculated. RMSE is the normalised mean of the squared difference between the observations and the predicted values and R 2 has been calculated with respect to the identity line. Moreover, we tested whether the regression lines between forecasted values and observation are significantly different from the identity line, since to obtain exactly the