Accurate prediction of multivariate time series (MTS) data is crucial for scientific assessment of possible geological disaster risks to power facilities. However

the existing deep learning models tend to focus too much on temporal information and ignore the complex dependencies between features when dealing with MTS data. To solve this problem

a model of feature-time block multi-layer perceptron(FTMLP) is proposed

which is designed for processing pervasive monitoring data and incorporates a feature-time module and a smoothed quadratic loss function. The feature-time module identifies key features through the feature module and performs frequency-domain filtering in time module to reveal the interaction between feature and time information. In addition

the smoothed quadratic loss function is used to optimize the predictive performance of the model

which makes it better to deal with noise and out-liers in the data. Experimental results on several disaster monitoring datasets with pervasive characteristics show that the FTMLP model outperforms the existing MLP-based models in terms of both prediction accuracy and robustness

especially when dealing with complex feature dependencies and noisy data

and can significantly improve the prediction performance. The proposed FTMLP model effectively improves the prediction accuracy of multivariate time series data in disaster monitoring by fusing the feature-time module and smoothing the quadratic loss function. The innovation of the model lies in its ability to capture the complex interactions of feature and time information simultaneously and enhance the robustness of the model by optimizing the loss function

which provides a new effective tool for early warning and risk assessment of geological disaster in power facilities.