Fine particulate matter (PM 2.5 ) is a complex air pollutant with numerous gaseous and solid constituents. PM 2.5 possesses a significant hazard due to its ability to penetrate deep into the lungs, corrode the alveolar wall, and impair lung functions. Modeling the non-linear and dynamic time series of daily PM 2.5 concentration remains a challenge. This study proposes a deep LSTM neural network to forecast accurate PM 2.5 concentration in the Kathmandu valley. Correlation analysis illustrates that dew, minimum ambient temperature, maximum ambient temperature, and pressure are strongly correlated with PM 2.5 concentration. Hence, five models are developed based on different input parameter combinations and are eventually evaluated to determine the best performing model. Model 2 with single-step prediction is the best performing deep LSTM model with RMSE of 13.04 μg/m 3 and MAE of 10.81 μg/m 3 . The SARIMA model applied to the univariate PM 2.5 data series illustrates the RMSE of 19.54 μg/m 3 and MAE of 15.21 μg/m 3 for the test data. Hence, the deep LSTM model with past PM 2.5 data and dew as inputs is recommended to predict future PM 2.5 concentration in the Kathmandu valley. The negative impact of PM 2.5 concentration on public health can be minimized with efficient forecasting.

中文翻译：

---

探索深度 LSTM 神经网络以使用尼泊尔加德满都谷地的气象参数预测每日 PM2.5 浓度

细颗粒物 (PM 2.5 ) 是一种复杂的空气污染物，具有多种气态和固态成分。PM 2.5 具有显着的危害，因为它能够深入肺部，腐蚀肺泡壁并损害肺功能。对每日 PM 2.5 浓度的非线性和动态时间序列建模仍然是一个挑战。本研究提出了一个深度 LSTM 神经网络来准确预测加德满都谷地的 PM 2.5 浓度。相关性分析表明，露水、最低环境温度、最高环境温度和压力与 PM 2.5 浓度密切相关。因此，根据不同的输入参数组合开发了五个模型，并最终进行评估以确定性能最佳的模型。具有单步预测的模型 2 是性能最好的深度 LSTM 模型，RMSE 为 13.04 μg/m 3 ，MAE 为 10.81 μg/m 3 。应用于单变量 PM 2.5 数据系列的 SARIMA 模型说明测试数据的 RMSE 为 19.54 μg/m 3 和 MAE 为 15.21 μg/m 3 。因此，建议使用以过去 PM 2.5 数据和露水作为输入的深度 LSTM 模型来预测加德满都谷地未来的 PM 2.5 浓度。PM 2.5 浓度对公共卫生的负面影响可以通过有效的预测最小化。