本論文中結合物聯網以及機器學習,完成了一套智慧灌溉系統雛形。過程中使用過往的降雨量紀錄以及過去天氣氣象紀錄來預測未來的降雨量,並且使用了物聯網之技術,做出可以自行判斷土壤溼度,達到智慧灌溉的目的。 機器學習的部分考量遞歸神經網路和長短記憶類神經網路的差別,最後選擇使用門閘遞迴單元(以降低運算成本)和遞歸神經網路的結合,輔以過去十年之氣候觀測資料的輸入,來達到降雨量預測之目的。 基本上,降雨量和許多因子相關,包括地區/地形、季節、季風、颱風與氣候變遷等等,這些都會造成建立有效降雨模型的困難,而構成值得探究進一步分析研究的課題,例如針對訓練資料量與驗證資料量的分配比例不同所造成對訓練以及驗證效果乃至預測精準度影響的探討,以找出適切的分配比例就是其中要項之一。 接著,使用 LORA-IOT無線技術與各個感測器做無線連接,並啟動自動灌溉系統使土壤適時維持某一必要濕度以上,並且基於自行預測的降雨量與土壤濕度(乃至植物的生長期程),自動決策當下的灌溉水量。
Abstract: In this thesis, based on the techniques of the Internet of Things and machine learning a prototype of a smart irrigation system is made. During the process, past rainfall records and past weather and meteorological records are used to predict future rainfall, and the connection via technique of the Internet of Things is utilized such that the system can make judgments on its own based on soil moisture and rainfall prediction to achieve the purpose of smart irrigation. For the machine learning we first consider the differences between recurrent neural networks and long-short-memory neural networks, then the idea of gate recursive unit associated with recurrent neural networks is utilized (to reduce computing costs), for which the climate observation data from the past decade are used as the input to achieve the goal of rainfall forecast. Basically, rainfall is related to many factors, including regions / topography, seasons, monsoons, typhoons and climate change, etc., which will cause difficulties in establishing effective rainfall models. These in turn constitute certain topics worthy of further analysis, such as the exploration about the impact of the ratio of the amount of training data over the amount of verification data upon the training and verification results, and even upon the prediction accuracy, aiming to find out the appropriate ratio. Finally, the LORA-IOT wireless technology is employed to establish a wireless connection (associated with lots of sensors), and to setup an automatic irrigation system to keep the soil moisture above a certain necessary value in a timely manner. And based on the home made rainfall prediction plus the soil moisture measured (even with the plant growth phases), the system would automatically decide the current irrigation water volume.