Monitoring The Performance Of An Underground Hydroponic Farm

By 2050 the world’s population is expected to increase by 3.5 billion, most of whom will live in cities, placing greater stress on infrastructure, resources and food security (Garnett, 2013). Urban agriculture is being increasingly investigated as an alternative local source of food, notably using hydroponics. Integrating urban farms within dense environments has the potential to utilise waste infrastructure and resources within cities and at the same time provide jobs and improve food security (Pons et al., 2015; Benis, Reinhart and Ferrao, 2017). This paper presents the unique case study of monitoring the development of a start-up hydroponic farm in derelict London tunnels, Growing Underground.

Defined sometimes as plant factories, integrated greenhouses, or vertical farming, hydroponic farms are becoming increasingly considered in our cities by councils, non-profits and commercial enterprises. Their potential in terms of Life Cycle Assessment has been examined by many researchers (Despommier, 2011; Specht et al., 2013). However, there is a lack of understanding of how well these systems can be integrated into existing infrastructure, and how their environment can be optimized in an efficient manner. Scientific models for plant growth are largely empirical and developed for certain types of commercial plants (Körner et al., 2007; van Straten and Henten, 2010), hence it is not possible to rely on models alone. This work presents a unique case study of monitoring both qualitative and quantitative data in an existing hydroponic farm in London. We describe the monitoring design and demonstrate how this data is used by the urban farm for two purposes: (a) to optimize their environment on a day to day basis, and (b) to design the future extensions of the urban farm.

This paper begins by describing the monitoring network in Growing Underground, as well as the challenges faced to design such a network in this environment. The data is then cross-analysed with each other by finding interpretative parameters to establish dependencies, and sub-dividing the dataset into interesting groups (based on seasonality, crop performance, optimal conditions). This leads to the final section, which presents examples of how the data is used to derive key performance indicators and track these indicators under changing conditions, in future designs of the farm.

Growing Underground is an unheated hydroponic farm, installed in tunnels designed as a WW2 air raid shelter in the 1940s, 33m below ground (Zero Carbon Foods Ltd, 2017). Founders Richard Ballard and Steven Dring wished to integrate hydroponics into derelict space in London, to grow microgreens closer to their point of use, which spurred the development of the farm in these tunnels abandoned since they last housed the Windrush generation of migrants in the 60s. The farm initially catered to hotels and restaurants and has expanded to become a supplier of M&S and Ocado. The main crops are peashoots, a variety of basil, coriander, parsley, salad rocket, pink radish and mustard plants. The farm consists of two parallel tunnels, running on two levels and spanning approximately 400 m. Currently only half the upper tunnel is used for farming, with a growing space of around 560 m², and is projected for major extensions in the near future.

Traditionally, hydroponic farms are glass greenhouses with controlled environments and high energy use intensity, thus this farm is a unique environment. Literature surrounding optimal growing conditions in greenhouses focus on maintaining temperature, relative humidity, air velocity and CO₂ (to some extent) within ideal bounds. Managing these fine bounds within a long tunnel environment, while optimising the use of energy, is a difficult task for the farm operators on a day to day basis. Designed in the 1940s, the ventilation system of these ex-shelters has to rival with the expectations of Controlled Environment Agriculture (CEA), but careful studying of the environmental conditions can help optimise the farm configuration and highlight the important factors to consider.

The main objectives of monitoring can be defined by these two points:

1. Keep growing: maintain adequate conditions for optimal crop growth at all times

2. Optimise energy use: minimise the resources put into the farm (water, fans, lighting) while optimising conditions for crop growth and maintain structural integrity.

As sensor networks become increasingly used in horticulture, it is important to transform the data collected into information to lead to better practice, especially in bespoke sites such as Growing Underground (GU). The challenges in data-driven models do not lie in data analysis alone. Sensors are known for their limitations which include low battery power, limited computational capability and small memory (Aqeel-ur-Rehman et al., 2014). Consequently, a range of methods were deployed to collect data in GU: from the design of a sensor network which sends data in real time to a viewing platform, and to manual reporting of control conditions within the farm.

Five environmental variables are continuously monitored in Growing Underground: temperature, relative humidity, CO₂ concentration, air velocity and light levels. For each sensor, the time since installation, location, and measurement interval are listed in Table 1, and are mapped in Figure 2. The location of the sensors linked to the Wireless Sensor Network (WSN) is illustrated in Figure 2a, while those which require manual downloading are in Figure 2b.

The WSN consists of the sensors transmitting data to Raspberry Pi loggers that are configured to process the data and send it to servers directly managed by the research team in a convenient format, and also store the data on SD cards when the wireless service goes down. A web platform was built to visualise in real-time the data acquisition. When Growing Underground engaged the services of Datch Systems Limited to provide enhanced data visualisation capabilities, it was relatively easy to add scripts to send our data to Datch Systems’ servers as well.

Temperature (1), relative humidity (1) and CO2 (1) are captured using AdvanticSys IAQM-THCO2 sensors. These are mains powered and connect to each other (via Wireless Modbus bridges) and to two Modbus/RTU dataloggers via a wireless mesh network using IEEE 802.15.4 radios (Advanticsys, 2017). These loggers originally sent data to a dashboard accessible on the subscription AdvanticSys website but were reconfigured to allow greater data accessibility. The original dataloggers were replaced by two Raspberry Pi single-board computers running Linux and fitted with USB-to-RS-485 adapters (Raspi 1 and 2 in Table 1), linked to the ADSL router directly with an Ethernet cable.

Temperature (2) and air velocity (1) are measured with four E+E Elektronik EE671 anemometers, added in November 2017. Available with either analogue or Modbus RTU outputs, the latter was chosen to connect the sensors to two additional Raspberry Pi computers using an RS-485 adapter. Each Raspberry Pi data logger is connected to two air velocity sensors (Figure 2a). Data is sent from Raspi 6 and 7 over WiFi to the local router. Additional WiFi access points were installed in the tunnel to allow greater flexibility in sensors placement.

Temperature (3) corresponds to the measurements from the TinyTag data loggers (Gemini Data Loggers, 2018), pictured in Figure 3 and locations shown in Figure 2(b). These versatile sensors are battery powered and waterproof, but the data needs to be manually downloaded off each sensor every three months using an inductive pad. This allows the sensors to be placed in areas not covered by a wireless network, such as outside, in empty tunnels and at several points along the farm which do not have electrical sockets. They provide information about the spatial variations of temperature along the tunnels.

Camera sensors: One goal of the project was to monitor plant growth. It was therefore decided to use simple USB webcams connected to several Raspberry Pi computers (Raspi 3-8), with four cameras per Pi (Figure 3). These were configured to take a photo every 15 minutes. It was hoped that these images could be analysed to provide an indication of the rate of leaf growth. These sensors were not a success. As production in the farm moves from rack to rack, the cameras needed to be repositioned to point at a new batch of product. The team were never able to get the necessary ‘buy-in’ from the farm staff, possibly due to concerns over privacy. No useful data was collected from the cameras and their use is not considered further in this paper.

Temperature (4), RH (2), CO₂ (2) and light (1a and 1b) are monitored with the new battery-powered wireless sensors designed by the team (Figure 4). The sensors create a low-power wireless mesh network with each other so that data can be sent even when out of range from the WiFi network, and are fitted with an SD card in case the mesh fails. The sensors used were a COZIR Non-Dispersive Infrared (NDIR) CO2 sensor from Gas Sensing Limited, a SHT21 temperature and humidity sensor from Sensirion (an I²C version of the SHT7x sensor used in the AdvanticSys sensors) and three different light sensors to measure the photosynthetically active radiation (PAR) in the tunnel: a digital OPT3001 and two analogue photodiodes. The sensors were connected to a deRFmega128 wireless microcontroller from Dresden Elektronik – based around the Atmel Atmega128RFA1 chip. The firmware for the new device was based on Contiki OS (Dunkels et al. 2004) using the 6LoWPAN/IPv6 radio stack. The datalogger consists of a Raspberry Pi connected to a deRFmega128 module with no sensors. It connects to the Internet via a WiFi connection to the ADSL router.

It was not possible to automatically log crop data and individual actions on the farm. We have therefore assembled a set of data that is manually recorded (described in Table 2). For instance, the specific location, type and growth period of every single crop is recorded into an online spreadsheet by the farm workers. This data gives information about what type of crop was being grown in the farm over time, as illustrated in Figure 5.

Qualitative comments about the farm’s environmental conditions are inferred from a text-based keyword analysis to mark periods of operational change. Trigger dates to look at contain comments such as: “dry/wet” crops, movement of equipment, doors were opened or closed, “lots of condensation” (example of comments in Figure 6).

Gathering data in an operating hydroponic farm presents challenges of metering in a high humidity environment with continuous operations in the farm (Teitel, Atias and Barak, 2010; Chaudhary, Nayse and Waghmare, 2011; Park and Park, 2011). In addition, an underground environment poses extra challenges of network connectivity and limited locations where sensors can be feasibly installed. As a result, a range of different sensors and recording methods were trialled over three years. For instance, energy readings were recorded manually for two years on week days until an automatic meter was installed. Regarding air velocity, several field visits taking point measurements with a thermal anemometer were necessary to identify the most representative locations to install the air velocity sensors. Some sensors could not connect to the wireless network, manual data was not recorded at weekends or during holidays, and there were periods when sensors were displaced or ran out of battery power. A major limitation to sensing was also the limited points of access to mains electrical power in the tunnels.

It is also important to use several different sensors, each monitoring the same parameter, to cross-calibrate the data. For instance, CO₂ concentration is measured through NDIR (non-dispersive infrared) which has a tendency to drift over time, and sensors typically auto-calibrate against the daily minimum, which is set to be atmospheric (400 ppm). As this is not the case on our site, the sensors require regular manual calibration with outside air, otherwise they can only be used to analyse the relative change in CO₂ levels, but not the absolute value.

To derive meaning from the measured data, the measurements were combined into a set of parameters which are described in Table 3. When daily data were not available (such as water tank levels, mass harvested or energy use), they were disaggregated to a daily level proportionally. Furthermore, temperature, air velocity, RH and crop growth data were available for different locations along the tunnel, which enabled interpretation of the environment at different locations within the tunnels and as well as across time.

The threshold values of optimal environmental conditions for the microgreens grown in the tunnel were first identified through a literature review, and then adapted through observation of the crop health over the three years of monitoring. These are summarised in Table 4. For example, relative humidity needs to be lower than conventional greenhouses due to condensation and the high risk of mould given the limited airflow in the tunnel. Temperature patterns depend on LED lights, which also serve as the main heating source in the tunnels. The monitoring data can thus be used to ensure these five environmental variables stay within the bounds indicated in Table 4.

We found that crop growth could be characterised by three variables that are not necessarily correlated: (1) number of growing days to reach desired size, (2) yield per area (for given number of seeds), and (3) crop health (percentage of crops diseased).

Crop yield data was divided into longitudinal zones along the tunnels (referred to as columns) in order to better understand how the spatial variations of environmental conditions affect crop growth. The data showed that, on an average, the crop yield was best in the middle of the farm (columns 9-20). The lower plot in Figure 7 shows for example that wasabi mustard tended to grow less well towards the end of the tunnels, in columns 30-31. The lower yields on the edges of the farm are attributed to the lower temperatures (Figure 9) and lower air velocities (Figure 12). Crops also tended to grow less well on the bottom shelf (colder temperatures), but there was no significant difference between the left and right aisles, suggesting sufficiently symmetrical conditions.

The number of growing days decreased when outdoor temperatures rose, whereas yields tended to be higher in winter. However, the data also suggests that each crop behaved differently and were not influenced by factors in the same way, but interpretations are limited as not all crops were grown in all conditions. At present, different crops are grown at positions along the tunnel based on available space and habit, rather than tailoring each crop location to their most suited environment (Figure 8). As a result, there is a bias in analysing location versus yield, as there are not as many datapoints for each crop at every location. This highlights the need for ongoing monitoring and regular data analysis in an operational farm, so that the impact of daily activities in the farm can be directly fed back to Growing Underground, and they can optimise the location where each crop is grown.

Weather conditions were found to affect temperature and humidity in the farm, energy use for the extraction fans and the number of growing days of the crops. Indeed, Figure 10 shows that as temperatures grew warmer over summer, the crops tended to grow faster to reach the same yields. At the same time, the overall yield is poorer in the summer, possibly due to it being too hot. Indeed, although it does get cooler than the lower bound of temperature in winter, crops tended to yield better, as the heat gained through the lighting system was enough to compensate for the cooler temperatures. This was not the case for the crops located near to the doors (columns 1-6 and 30-31 in Figure 7).

Because of this analysis, the crop yields were much better in summer 2018 than in the previous two summers, possibly due to better environmental management, with increased use of the extraction fan, carefully management of the doors for airflow and improved fan configurations (see next section).

The monitoring data was also used to educate operational changes in the farm. For example, the high levels of relative humidity led to the installation of dehumidifiers. For example, adding perforated tubes over the plant trays increased the air velocity at the front of the tunnel (Figure 12a).

Figure 11 shows how the relative humidity was within the optimal range for longer periods in 2018 The black points show the average outdoor monthly humidity, which shows little variation year on year, but clearly influences the stability of the relative humidity in the farm across winter and summer. Even with decreased ventilation (lower ACH needed to minimise cold air intake) in winter, the relative humidity levels are higher in the farm than in summer.

Different configurations of fans were tested, to increase air mixing in the farm, such as snail fans along the sides, perforated tubes, central fans, and experimenting with the extraction fan levels. The ultimate configuration led to lesser horizontal and vertical variation of temperature. and more optimal air velocities around the farm (Figure 12b). This led to more even crop growth longitudinally along the tunnel, and higher CO₂ uptake from the air. For example, adding perforated tubes over the plant trays increased the air velocity at the front of the tunnel (Figure 12a).

On average, the farm uses 56 kWh a day, and has two points of connection to the grid, at either end of the tunnel, known as Carpenters Place (CP) and Clapham Common (CC). The meter reading at CP can be used to derive the energy use for the first extraction fan, the baseload and the lighting. The second meter reading from CC is only linked to the second extraction fan.

The data showed that the temperatures in the tunnel were strongly influenced by the waste heat gained through the lighting (shown in Figure 13). This information was used to advise GU on the air changes per hour (ACH) with the extraction fans, and optimizing their lighting period, not just for longer photo-period, but for maintaining optimal temperatures. The duration of the photo-period impacts farm temperatures, which need to kept low in summer, but higher in winter. Farm conditions are also strongly influenced by the extraction fan, where augmenting the cross-ventilation will increase crop growth in summer but decrease it in winter.

This data analysis demonstrates the first exercise in using the monitored data to assist Growing Underground to fine tune the environment of the farm. Different types of data collection were discussed: daily manual recording and wireless sensor networks, for three types of parameters: crop growth, environmental conditions, and operational controls. The data was combined and compared with each other to derive meaning and establish optimal environmental conditions, ventilation configurations, and understanding the effects of seasonality. This work has thus shown how monitoring can be developed for an urban farm in a derelict space, but also how performance indicators can be drawn out from different kinds of data, in order to assist the growth and limit the energy use in a hydroponic farm.

The longitudinal and ongoing data collection lends itself quite well to deriving data-centric models to forecast controls within the farm. There is a growing trend to fully automate greenhouse controls (Chaudhary, Nayse and Waghmare, 2011; Körner and Hansen, 2012; Kozai, 2016), however these systems are still in their infancy and require adept statistical models based on extended monitoring. The interpretations established in this paper will be developed further in a forecasting model to test future farm configurations for extension projects.

We thank the cooperation of Growing Underground with our work, and for enabling such a fruitful collaboration for so many years. Project funded by the EPSRC.