Comparison performance analysis of autoregressive integrated moving average and deep learning long-short term memory forecasting weather data
In one sentence. LSTM beats ARIMA on most station weather variables such as temperature, humidity and radiation, and in computation time, though precipitation errors remain highest.
Abstract
Information about the weather is crucial in assisting human activities and labor because the weather is a factor that cannot be separated and is closely related to all human activities. The purpose this study to compare performance the Autoregressive Integrated Moving Average (AIMA) and Long-Short Term Memory (LSTM) algorithm models with case studies of weather forecasting. This study uses comparison of two methods, forecasting using AIMA and LSTM methods. LSTM method provides the best forecasting performance for attribute minimum temperature, maximum temperature, and average temperature with the Root mean squared error value below 1.45 and the Mean Absolute Error value below 1.14. For attributes of average humidity and solar radiation with a Root mean squared error value of 2.62 to 3.82 and a Mean Absolute Error value of 2.21 to 3.2. Precipitation forecasting has the highest error value with a root mean squared error value of 9.99 and a mean absolute error of 6.5. The AIMA method provides the best forecasting performance on the attribute minimum temperature, maximum temperature, and average temperature with the Root mean squared error value below 1.47 and the Mean Absolute Error value below 1.16. For the sun exposure attribute with a Root mean squared error value of 2.91 to 3.05. Whereas the average humidity attribute has the highest error with the Root mean squared error value reaching 4.97 and the Mean Absolute Error reaching 3.99. LSTM method is better in terms of forecasting results and in terms of computation time. From every forecast made, the LSTM method produces a smaller error value.
Labels
- Topic
- Local & station forecasting
- Type
- Evaluation or intercomparison of existing models
- Method
- RNN / LSTM / ConvLSTM, Classical ML (trees, GP, linear, kernels)
- Time scale
- 6 h–3 days
- Domain
- Local / station
- Data
- Station / in-situ
- Status
- Research
- Identifiers
- DOI:10.54076/jumpa.v3i1.302 OpenAlex:W4365151793
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.