Using machine learning techniques to predict liquid dairy manure temperature during storage

Abstract There is no standard method to predict manure temperature during storage. So, decision support tools, on-farm nutrient cycling models, and life cycle assessment tools to assess the sustainability of agricultural production systems that include manure typically use ambient air temperature as a surrogate for manure temperature. This study explores the application of machine learning algorithms' unique abilities to predict manure temperature based on measured data. The data was collected from two on-farm manure storages (clay pit and concrete tank) instrumented with sensors to acquire manure temperature at various depths during the storage period. The local weather data (ambient air temperature, wind speed, wind direction, solar radiation, relative humidity, and rainfall) were recorded by stations installed at each farm. The data were subjected to four machine learning algorithms gradient boosted trees, bagged tree ensembles, random forest ensembles, and neural networks using the supervised learning approach. The weather data and two additional parameters, time (month) and the manure depth above a sensor, were derived and used as inputs for the machine learning algorithms. Further, the developed machine learning algorithms were challenged with parameters from historical weather data (1990 to 2020) to assess their suitability to predict manure temperature where local weather is not available. The results showed that, in general, the stored manure temperature lagged but followed a similar trend as the ambient air temperature and solar radiation. The average manure temperature was higher than the ambient air temperature for most of the year. Depth influenced the manure temperature; manure in the top layers had a higher temperature during warm periods than the bottom layers, and vice versa during cold seasons. The ensemble models performed better than the neural networks by predicting manure temperatures closer to the measured values and predictions during the scenario analysis. The random forests and bagged tree ensembles were the best performers. Models tended to make better predictions as the depth of manure above a sensor increased. This work will provide added value for developing better decision support tools and models for assessing nutrient cycling on farms. It also informs our knowledge to develop emission mitigation strategies during manure storage, leading to more sustainable manure management practices.