We're exploring a solution for prediction of overheat in room temperatures. We'd like to learn from your experience and recommendations of methods or models. We have collected time series data of the room temperatures at the sites. The sampling period is 15 minutes. We would like to predict whether there would be an overheat in one hour. The overheat event is defined by the temperature at a room is above 95 F. (Note, our requirements do not insist a regression but a classification would suffice.) The rooms house industry computers, which generate much heat. Usually, air-conditioners are installed in the rooms. When the air-conditioners stop working, or the windows are opened, the overheat may happen. We have tried a few solutions. 1. ARIMA 2. Fourier analysis 3. A liner model of f(t+4) = f(t) + rolling_average(f(t))*4 4. A encoder-decoder neural network to recognize the rising edge of the time series All the above attempts only works for some cases. We're not sure if more fine tuning is needed, or we need to try more appropriate models? Please share your insights and wisdom. Thanks in advance!

Time series modeling works best when you've got observable patterns in your target (e.g. temperature) and a sufficiently long period of time to model the periodic trends. Have you graphed the data to see if there is a pattern? Additionally, you should be capturing exogenous variables - AC not working, window opened, outside temperature - as those are probably strong indicators of when you'd overheat. And thinking about how to predict those as inputs to an algorithm.

Thanks for the discussion! We have indoor temperature measurements for the last two and half years. There is apparent cycle pattern of daily 24 hours. As we discussed, we just don't have the data of AC not working, or windowed opened. We do have some outdoor temperatures, and we may start to collect more. So far, we found that a linear model of A liner model of f(t+4) = f(t) + rolling_average(f(t))*4 tracks very well with the past overheat history. But the model does not work well for general future prediction, as it overshoots significantly when the temperatures oscillate rapidly. (That is the model has good recall, but may have poor precision as far as temperature regression is concerned.) Currently, we may use the linear model as a base for overheat perdition together in considerations of the other necessary conditions, such as outdoor temperatures, the current temperatures, and the past history of the sites being overheat or not. I'd like to learn if there are better methodologies or workflows, to make the model more robust, and simpler.

Statsforecast's AutoARIMA will iterate through rolling average, lag and difference windows for you. That would determine if something other than 4 is a better value. In addition, because you have at daily seasonality you may want to consider using MSTL + AutoARIMA with daily and weekly trends: https://nixtlaverse.nixtla.io/statsforecast/docs/models/multipleseasonaltrend.html Regarding exogenous variables I would consider that should be possible to get historical and predict: 1. outside temperature 2. duration since A/C was last serviced (a good proxy might be since last outage) 3. power consumption - kilowatts per hour

Very good points! I'll try AutoARIMA, and MSTL + AutoARIMA What is an example way of considering exogenous? How to do it with AutoARIMA, for example? Your points of considering the possibility of sensor problems and power consumption might be interesting. So far, I observed that the temperature oscillates, only some time period of a day, typically, in the evening or night. I suspect that it might be the dynamic interaction of AC and heating. I'll look into if there is a problem of bad temperature sensors.