Short-term forecast improvement of maximum temperature by state-space model approach: the study case of the TO CHAIR project

Abstract

In the context of ''TO CHAIR'' project, this work aims to improve the accuracy of short-term forecasts of maximum air temperature obtained from the https://weatherstack.com/ website. The proposed methodology is based on a state-space representation that incorporates the latent process, the state, which is estimated recursively using the Kalman filter. The proposed model linearly and stochastically relates the forecasts from the website (as a covariate) to the observations of the maximum temperature recorded at the study site. The specification of the state-space model is performed using the maximum likelihood method under the assumption of normality of errors, where empirical confidence intervals are presented. In addition, this work also presents a treatment of outliers based on the ratios between the observed maximum temperature and the website forecasts.

Comparison of outlier detection approaches in a Smart Cities sensor data context

“…In comparison with other studies, the outliers occurrence here is similar to that that has been reported in other studies, both in "smart-cities" contexts (Csaji et A notable point of our work is the identification of the need for data filtering on temperature and PM data before outlier detection to omit extreme data. On temperature data, the filter should be adapted to the city's climate conditions (see also Pereira et al, 2023;Ma et al, 2017). PM should be eliminated at a range of 0-1,000 μm/m 3 , so extreme values can be excluded (see also Aix et One notable example is the presence of environmental phenomena that can be overshadowed by outliers and the methods used for their estimation and management.…”

Improving Predictive Accuracy in the Context of Dynamic Modelling of Non-Stationary Time Series with Outliers

“…State-space models were originally developed in aerospace engineering in the early 1960s for the purpose of monitoring and correcting the trajectory of a spacecraft headed to the moon. Today, these models have wide applicability in many areas, such as finances [1], ecology [2], machine learning [3], and time series analysis and forecasting [4][5][6][7]. These models, associated with the Kalman filter algorithm [8], are a very powerful tool given their ability to update predictions both in real time and in a recursive procedure as new observations of the time series become available, thus improving the accuracy of predictions.…”