以機器學習技術建置溫室智慧噴霧系統

Build an intelligent greenhouse spraying-system using machine learning techniques

陳廷軒(Ting-Hsuan Chen)

指導教授 ： 張斐章

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102542

摘要

溫室耕種最主要的優勢係藉由建築特性及環控策略，以維持內部環境穩定，達到產量最佳化，然而其相較於傳統露地栽培亦會消耗更多資源，因此須以水-能源-糧食鏈結(Water-Energy-Food Nexus)之管理為出發點，對不同環控策略進行量化與效益分析，方能妥善利用資源。為此，本研究先以質量守恆與能量守恆兩個公式，建置一套能推估下一小時溫室內溫度與相對濕度的物理模式，用以模擬噴霧後溫室內之環境，並以彰化縣伸港鄉溫室內物聯網設備收集之歷史監測資料，驗證物理模式的準確度與可靠度。其驗證流程為在N=0時(N代表單位為一小時之時距間隔)，以第一筆資料中的溫室內溫度與相對濕度，當作模式ｔ時刻溫室內溫度與相對濕度的初始值，再引入t+N時刻之溫室外溫度、溫室外相對濕度、溫室外日照量、風速與風向，推估t+N+1時刻的溫室內溫度與相對濕度，最後將此t+N+1的溫室內溫度與相對溼度，引入物理模式做為推進至下一時刻(N=N+1)的溫室內溫度與相對溼度起始值，重複此步驟直至觀測資料全數跑完。分析結果顯示了，物理模式在推估下一小時溫室內溫度與相對濕度的決定係數值為0.79及0.80，均方根誤差值則為1.89℃及8.17%，這代表物理模式在推估溫室內溫度與相對溼度的變化上，具有良好的精確度與可靠度。

接著建置一套能預測下一小時溫室內溫度與相對濕度的類神經網路(ANN)預測模式，並以同一批監測資料驗證模式的準確度與可靠度。其驗證流程為在N=0時，以t+N時刻的溫室外溫度、溫室內溫度、溫室外相對濕度、溫室內相對濕度、溫室外日照量以及風速等六項因子作為輸入因子，引入倒傳遞類神經網路預測t+N+1時刻的溫室內溫度與相對溼度，完成之後推進至下一時刻(N=N+1)，重複此步驟直至觀測資料全數跑完。分析結果顯示了，ANN預測模式在預測下一小時溫室內溫度與相對濕度的決定係數值R2可達0.82及0.88，均方根誤差值RMSE則為1.55℃及4.19%，這顯示了預測模式在預測溫室內溫度與相對溼度的變化上，具有良好的的精確度與可靠度，且表現優於物理模式。而因本研究因屬研發性質，尚無各時刻之噴霧量、噴霧後溫室內溫度與相對溼度的實際監測資料，導致預測模式無法直接計算出各時刻所需噴霧量，以及噴霧後溫室內溫度與相對溼度，故需將預測模式與物理模式結合，先以預測模式得出下一小時溫室內溫度與相對溼度的預測值後，再以物理模式計算出能將預測值改變至適合植物生長環境所需的噴霧量，以及噴霧後的溫室內溫度與相對溼度。最後一步才會是探討智慧噴霧系統與傳統噴霧方式的噴霧效果，以及噴霧前後的資源消耗差異。

本研究以彰化縣伸港鄉占地面積1560平方公尺的強固型開頂溫室，時間範圍為2019年5月20日00:00至2019年7月20日23:00共1488筆之溫室歷史監測資料，模擬溫室內種植番茄時所需之噴霧方式。根據研究結果顯示，傳統噴霧方式與智慧噴霧系統皆對溫室有良好的環控能力，惟傳統噴霧方式需消耗129478公斤的水及90度電，而智慧噴霧系統則只需42962公斤的水及29.8度電，省水率及省電率均高達66.8%; 此一結果顯示本研究所發展的智慧噴霧系統，有助於溫室內環境的自動控制且能省水及省電，具有極大的應用效益，同時亦補強了大多數溫室環境自動控制的相關研究，對於資源消耗著墨過少的問題。

關鍵字

水-能源-糧食鏈結 ； 溫室 ； 機器學習 ； 物聯網

並列摘要

The main advantage of greenhouse farming is to maintain a stable internal environment suitable for crop growing to achieve optimal yields through architectural characteristics and environmental control strategies. Greenhouse farming depletes more resources than open field cultivation. Therefore we should take Water-Food-Energy Nexus (WFE Nexus) management as a starting point to increase resources utilization efficiency.

For this purpose, we first construct a physical model to predict one-hour-ahead indoor temperature and relative humidity of a greenhouse by the conservation of mass equation and the conservation of energy equation. The predicted value is used to simulate the environment of the greenhouse after spraying. In the second step, we validate the accuracy and reliability of the physical model with the real data recorded by ‘Internet of Things’ (IoT) devices in a greenhouse located in Changhua County, Taiwan. The validation process takes the indoor temperature and relative humidity of the first time series data as initial values when N=0 (N represents the time step at an hourly scale). Then we incorporate outdoor temperature, outdoor relative humidity, outdoor solar radiation, wind speed, and wind direction at t+N into the physical model. Therefore, the physical model is able to provide indoor temperature and relative humidity at t+N+1. Finally, we feed the physical model with these outputs as the initial values of the next time-step (N=N+1). This validation process continues until all the time-steps of monitored data are finished. According to the analysis results, the coefficient of determination is 0.79 for temperature and 0.80 for relative humidity, while the root-mean- square error is 1.89℃ for temperature and 8.17% for relative humidity. These results show that the physical model provides good accuracy and reliability.

We also build a prediction model by using the back-propagation neural network (BPNN) to predict one-hour-ahead indoor temperature and relative humidity. Then we validate the accuracy and reliability of the machine learning-based model with monitored data. The validation process incorporates indoor temperature, indoor relative humidity, outdoor temperature, outdoor relative humidity, outdoor solar radiation, and wind speed at t+N into the prediction model. Therefore, the prediction model is able to output indoor temperature and relative humidity at t+N+1, and then moves to the next time-step(N=N+1).This validation process continues until all the time-steps of monitored data are finished. According to the analysis results, the coefficient of determination is 0.82 for temperature and 0.88 for relative humidity, while the root-mean- square error is 1.55℃ for temperature and 4.19% for relative humidity. These results show that the prediction model not only provides good accuracy and reliability but also performs a little bit better than the physical model.

Because this research belongs to research and development (R D) processes, data of spraying amount, and the temperature and relative humidity after spraying are unavailable. Therefore, the prediction model is unable to directly calculate the spraying amount, or the temperature and relative humidity after spraying. In order to solve this problem, we combine the physical model and the prediction model into an intelligent spraying-system for greenhouse farming. The prediction model is used for forecasting indoor temperature and relative humidity at t+N+1. The physical model is used for simulating the environment of greenhouses after spraying and calculating the spraying amount required to maintain a stable internal environment. The final step is to explore the effects of the traditional spraying method and the proposed intelligent spraying-system as well as their consumption of resources before and after spraying.

The data for use in this research were collected by IoT devices in a greenhouse located in Changhua County, Taiwan from 00:00 May 20th, 2019 to 23:00 July 20th, 2019. The investigative greenhouse is an open-top greenhouse and occupies an area of about 1560 square meters. We use the monitored data of the greenhouse to simulate the spraying pattern for growing tomatoes. According to the analysis results, both the traditional spraying method and the proposed intelligent spraying-system can well maintain the environment of the greenhouse. The traditional spraying method consumes 129477.9 kg of water and 90 KWh of electricity. In contrast, the proposed intelligent spraying-system only consumes 42961.6 kg of water and 29.8 KWh of electricity. It is worth noting that the proposed intelligent spraying-system can save about 66.8% of water and energy, compared to the traditional spraying method. The results show that the proposed intelligent spraying-system not only conduce to the automatic control of greenhouse environment but also saves resources. No doubt the proposed intelligent spraying-system provides great applicability and reinforces the issue that most researches on the automatic control of greenhouse environment rarely address the consumption of resources.

並列關鍵字

Water-Food-Energy Nexus (WFE Nexus) ； Greenhouse ； Machine learning ； Internet of Things (IoT)

參考文獻

1. Al-Mezeini, N. K., Oukil, A., Al-Ismaili, A. M. (2020). Investigating the efficiency of greenhouse production in Oman: A two-stage approach based on Data Envelopment Analysis and double bootstrapping. Journal of Cleaner Production, 247, 119160.

2. Ammar, M. E., Davies, E. G. (2019). On the accuracy of crop production and water requirement calculations: Process-based crop modeling at daily, semi-weekly, and weekly time steps for integrated assessments. Journal of environmental management, 238, 460-472.

3. Bottcher, R. W., Baughman, G. R., Gates, R. S., Timmons, M. B. (1991). Characterizing erficiency of misting systems for poultry. Transactions of the ASAE, 34(2), 586-0590.

4. Chang, F. J., Tsai, M. J. (2016). A nonlinear spatio-temporal lumping of radar rainfall for modeling multi-step-ahead inflow forecasts by data-driven techniques. Journal of Hydrology, 535, 256-269.

5. Chang, F. J., Chang, L. C., Kao, H. S., Wu, G. R. (2010). Assessing the effort of meteorological variables for evaporation estimation by self-organizing map neural network. Journal of Hydrology, 384(1-2), 118-129.