卷積層與倒傳遞混合式類神經網路針對大區域之PM2.5進行長時距預測

近來隨著人們越來越關心健康惡化的問題，細懸浮微粒(PM2.5)越發的得到社會上的關注。建立精準的PM2.5預測模式仍面臨著巨大的挑戰因爲對於它動態的化學反應且排放數據及預測的不確定性。本研究將會建立一個混合模型 (CNN-BP) ，它是使用卷積神經網絡（CNN）和倒傳播神經網絡（BPNN）來同時對多個測站進行準確的長時距PM2.5預測。此研究用了2017年的73個空品測站的資料，并且用了六個空污因子以及兩個氣象因子為輸入模式輸入。總共有639,480筆資料并切分成訓練(409,238,64%)，驗證(102,346, 16%)以及測試 (127,896, 20%) 階段。PM2.5的預測被視爲是以空污因子及七項因子為參數之函數。此研究提出的CNN-BP模式有效地學習到輸入數據的主要特徵，並同時預測了區域性之PM2.5長時距預測 (73個測站；t+1-t+10)。本研究結果反映出區域性長時距的預測，利用混合式CNN-BP可獲得比BPNN, 隨機森林(random forest)及長短期記憶體類神經網路(LSTM)要高的準確率及卓越的表現。此外，CNN-BP模式不但可藉由高維度異質性的長延滯輸入資料來應對因維度被降低而導致重要資訊遺失的問題；也因此有著探索全臺灣五個區域(R1-R5)不同機制(本地排放和遠距離跨界傳播)的能力。此研究證明多個測站(區域性) 之長時距預測問題是可以用一個模式(CNN-BP)來完成這個目的，達到一個新的里程碑。因此，CNN-BP可以促進即時的PM2.5預報服務，并且可以在綫公開提供預報。

關鍵字

PM2.5預測；深度學習；卷積層類神經網路；倒傳遞類神經網路；長時距之預測；臺灣

並列摘要

It is well known that some diseases are caused by fine particulate matters such as PM2.5. Therefore, the forecast of such particles gains an increasing public concern. Due to the uncertainty of the dynamic formulation process of the tiny particles (emission data and their projections), modelling PM2.5 concentrations remains a challenging task. Since PM2.5 forecasting is a complex non-linear problem, artificial neural networks (ANNs) are often used in previous studies. The hybrid model (CNN-BP), which is a fusion of the Convolutional Neural Network (CNN) and a Back Propagation Neural Network (BPNN), is proposed in this study to make accurate PM2.5 forecasts for multiple stations at multiple horizons simultaneously through effectively learning the dominant features of input data. The datasets used for the PM2.5 forecasting in this study are related to six air quality and two meteorological factors (hourly) collected from 73 air quality monitoring stations in Taiwan during 2017. The hourly monitoring datasets (639,480) were allocated into training (409,238, 64%), validation (102,346, 16%), and testing (127,896, 20%) stages. The results demonstrate this proposed CNN-BP model can produce more accurate regional multi-step-ahead PM2.5 forecasts (73 stations; t+1−t+10), as compared to the widely used ANNs like BPNN, random forest and long short term memory neural network. With the ability to handle the heterogeneous inputs (with relatively large time-lags) adequately, the CNN-BP can mitigate the loss caused by dimension reduction. In addition, this hybrid model can explore the differences in PM2.5 mechanisms (local emission and transboundary transmission) for the 5 regions (R1-R5) and the whole Taiwan. In this study, exactly one model (i.e. the proposed CNN-BP model) is used to make multi-site (regional) and multi-horizon forecasting, which is a novel achievement. Consequently, this work suggests a real-time PM2.5 forecast service for the public.

並列關鍵字

PM2.5 forecast ； Deep learning ； Convolutional neural network ； Multistep ahead forecasts ； Artificial neural networks

參考文獻

Ausati, S., Amanollahi, J. (2016). Assessing the accuracy of ANFIS, EEMD-GRNN, PCR, and MLR models in predicting PM2. 5. Atmospheric environment, 142, 465-474. https://doi.org/10.1016/j.atmosenv.2016.08.007

Google Scholar

Bai, Y., Zeng, B., Li, C., Zhang, J. (2019). An ensemble long short-term memory neural network for hourly PM2. 5 concentration forecasting. Chemosphere, 222, 286-294. https://doi.org/10.1016/j.chemosphere.2019.01.121

Google Scholar

Borovykh, A., Bohte, S., Oosterlee, C. W. (2017). Conditional time series forecasting with convolutional neural networks. arXiv preprint arXiv:1703.04691.

Google Scholar

Chan, C. C., Chuang, K. J., Chien, L. C., Chen, W. J., Chang, W. T. (2006). Urban air pollution and emergency admissions for cerebrovascular diseases in Taipei, Taiwan. European heart journal, 27(10), 1238-1244. https://doi.org/10.1093/eurheartj/ehi835

Google Scholar

Chang, F. J., Chen, P. A., Lu, Y. R., Huang, E., Chang, K. Y. (2014). Real-time multi-step-ahead water level forecasting by recurrent neural networks for urban flood control. Journal of Hydrology, 517, 836-846. https://doi.org/10.1016/j.jhydrol.2014.06.013

Google Scholar

國際替代計量

卷積層與倒傳遞混合式類神經網路針對大區域之PM2.5進行長時距預測

全文下載

主題瀏覽