Temperature Forecasting on Low‐Cost Sensor Nodes: What a Public Benchmark Does Not Tell You

ABSTRACT Short‐term temperature forecasts are useful in many IoT systems, but the models behind them are usually tested on one large public dataset rather than on the cheap sensors actually deployed. We ran one forecasting pipeline—a naive “no change” baseline and three machine‐learning models (Ridge regression, Random Forest, and XGBoost) over lagged and rolling features—on two sources: the Jena Climate benchmark (Germany, 2009–2016, about 420 000 records) and 1 month of data from each of three low‐cost optical particle‐counter nodes in Tetovo, North Macedonia. All splits were chronological, and every evaluation was repeated with walk‐forward cross‐validation. Two of the three field nodes came close to the benchmark in absolute error (RMSE 0.76°C and 0.78°C, against 0.63°C on Jena) on roughly one hundredth of the training data, and two exported two‐feature models ran in real time on an Arduino MKR1000. The rest of the results were less comfortable. At one of those two nodes, no trained model beat the naive baseline, and walk‐forward evaluation showed the tree ensembles there to be far more unstable than the single split suggested. The third node's much larger error was not a modeling failure: its housing was heating in the sun, which we traced in the raw files and the site operators confirmed. Two further data‐quality defects, one per dataset, each changed a reported error figure. On this evidence, a benchmark score on its own says rather little about how a forecasting pipeline will behave on cheap edge hardware in the field.