TimeADF: A Predictive Method for Cross-Level Governmental Data Fusion
Xinyu Bi, Ziwei Yan, Junyu Zhu, Enguang Zuo · 2025
At present, the government service platform of Xinjiang Uygur Autonomous Region has key problems such as difficult business collaboration, cross-level, cross-field and cross-departmental information transmission, and processing and storage of multi-source heterogeneous information data. In response to these challenges, this paper proposes a construction scheme of the smart government affairs big data platform based on the Adaptive Decomposition and Filtering Framework (TimeADF). TimeADF solves the problems of high-frequency detail capture and long-term trend prediction of multi-source heterogeneous data through a three-stage design of reversible feature Reconstruction (RevIN), Frequency decomposition module (FDB), and high-frequency selector (ChooseK). Specifically, first normalize the multi-scale data using RevIN to ensure the integrity of cross-scale information; Then, through the Fourier transform, the low-frequency and medium-frequency components are separated by the FDB module, and the high-frequency components are calculated by the residuals, thereby dynamically screening the low-frequency policy cycle and high-frequency emergencies. The first k hidden high-frequency signals are extracted by the ChooseK module, and the hidden high-frequency information in the medium and low-frequency components is extracted for the second time. Because high-frequency information is more likely to predict the trends of certain government data in the future and enhance the sensitivity of prediction. Therefore, after superimposing the two high-frequency components and inputting them into FFN, the predictive output is achieved, accurately capturing the high-frequency details and long-term trends of government data. We conducted model checking on six datasets related to electricity, weather, financial transactions, etc. The prediction lengths covered {96, 192, 336, 720}. The results showed that TimeADF increased by an average of 39.8% and 24.7% respectively on MSE and MAE. It is significantly superior to prediction models such as MICN, DLinear, and Transformer.