When no optical feedback system is used, the solar tracker is operating in dead-reckoning or passive mode. In this mode, the sun's position is continuously calculated, and the tracker mirrors are adjusted to direct a stable image of the sun into the laboratory below. Total pointing accuracy in this mode is an accumulation of the many sources of error of the complete design. These include the following errors, listed in approximate increasing order of significance.

We chose to use a different solution for our optical feedback design: edge detection using four silicon diodes spaced evenly around the perimeter of a focused solar image of approximately 60 mm in diameter. This method was chosen primarily because the major sources of pointing errors are uncertainty in alignment and poor levelling. By their very nature, these errors appear as a slow-changing function, often sinusoidal, with a wavelength measured in hours. We do not want information from the bright solar image itself because there is a possibility that this could make the system too sensitive to rapid intensity changes caused by passing clouds. The real information instead comes from the high contrast of the intensity gradient found at the perimeter of the solar image.

The size of our focused image is large compared to that used with a typical quadrant detector. The larger image requires longer optics (for the same focal length lens). Each of our installations is different. We have used a variety of solar image sizes ranging from 20 to 60 mm, depending upon the lenses available at the time and the size and location of the free space available to mount the optics. The edge detection diodes are nominally named left, right, up, and down (L, R, U, and D), with an active surface area chosen to suit the solar image size. In our examples we used diodes with about 7 mm2. The diodes are positioned to be partially illuminated by the edge of the image, thus ensuring a high response to any movement of the image across their surface. The diode signals are lower in level and exhibit more noise compared to those from a quadrant detector. This is partially due to the low intensity of the solar image. They are also sensitive to image “jitter” caused by atmospheric turbulence and by constant movement of the mirrors by the tracker itself. If necessary, diode signals can be improved by increasing the size of the sample mirror – but at the expense of using some of the light available for the spectrometer. Diode signals are further improved with an adjustable gain operational amplifier circuit near the diodes, and the jitter is somewhat smoothed in software using a running average of 10 samples.

Despite the differences in solar image size, intensity, and sample pick-off size, each installation appears to perform similarly. This indicates that the edge detection concept is robust in response to the various trade-offs in design parameters.

Figure 3 shows an example of our feedback optics. Firstly, a small (∼ 2 mm2) front-surface mirror (Fig. 3a) samples the main solar beam from the tracker above. This is aimed at a distant wall as a useful visual reference for the operator to quickly assess tracker performance or look for the presence of cloud (Fig. 3b). A slightly larger non-vignetted portion of the main beam is directed into the feedback optics using a small mirror with dimensions of about 10 mm × 10 mm (Fig. 3c), with adjustment performed by the alignment mirror (Fig. 3d). This sample is focused with a small spotting telescope, half a set of binoculars (Fig. 3e), or a custom objective lens and eyepiece combination. The image is focused onto the diodes and amplifier printed circuit board (PCB); see Fig. 3g.

Focus and positioning can be checked by placing white paper in front of the diodes. The image should be centred, circular, and sharp. On the larger images, major sunspots may be seen. Electronic gain is adjusted to equal signal levels from the four channels, with emphasis on getting the levels in each pair equal (i.e. L = R, U = D). Physical protection and shielding from stray light are achieved using a sheet of inexpensive plastic IR filter (Optolite™ Industrial Plastics, UK) just in front of the diodes (Fig. 3g). The photodiode chosen needs a wavelength response in the near IR that matches any external filter used, e.g. PIN diode type BPW41N or BPW34F (Vishay). The signals are then routed to the tracker electronics box and sampled at 1 to 2 Hz using four channels of the 10-bit analogue to digital convertor (ADC) within our motor control PCB. The algorithms that translates these values into motor movements are described below in Sect. 3.2 and 3.3.

Two parameters, set in a configuration file, are used to control decision making within the optical feedback system. The “threshold” parameter is used to set the solar intensity level needed to initiate the use of optical feedback. During initial tracker installation, this threshold parameter is set to be somewhat below the minimum expected clear-sky diode signal level. This makes some allowance for taking measurements through haze or thin cloud (if required), and for the effect of mirrors becoming dirty over time. Each (averaged) diode signal level is compared to the threshold parameter value. If no diode signals exceed this value, then the sky is deemed too cloudy and optical feedback is not engaged. The tracker continues following the sun using passive mode. If one or more diodes exceed the threshold value, then optical feedback is activated.

When optical feedback is active, averaged L–R and U–D pairs are analysed individually using a “hysteresis” parameter. The signal difference within each pair must be greater than this value before a control action is taken. The hysteresis parameter helps prevent unnecessary control actions (or “hunting”). If the signal difference within a pair exceeds the hysteresis parameter value, the algorithms act to steer the mirrors to minimise this difference by incrementing or decrementing a combination of azimuth and elevation error offset variables by 0.001∘ about every second. These error offsets are then used as a reference for the azimuth or elevation rotator positioning. The hysteresis parameter works well if set at about 10 %–30 % of the threshold value and seems to have little effect on accuracy if the image perimeter and diode positions are well matched. Movement priority is given to correct the diode pair with the greatest difference in intensities. If all four signals satisfy hysteresis and threshold requirements the tracker can be called “locked”, and no further corrective feedback is necessary. On a well-tuned system with clear sky, the tracker can be locked for longer than a minute. Fast passing clouds have little effect on tracker pointing due to the long time constant of the feedback loop. Any significant cloudiness is recorded in a logged file for post-processing of measurements and can also be used to halt automated observations until the sky is clear again. When the sky is too cloudy, the tracker reverts to passive mode, and continues tracking the calculated position of the sun.

At this stage it is important to note that there is no simple relationship between the required movement to equalise signals in a diode pair (e.g. U and D) and a single error offset variable (e.g. elevation). In other words, we will not necessarily move up and down at the reference plane by solely moving the elevation mirror. This is because the image at the diode reference plane rotates as the day progresses and is additionally modified by translation of the image through the feedback optics. This process is further explained below in Sect. 3.3.

An algorithm is needed to map the required image movement directions, relative to the feedback plane (i.e. up, down, left, right), into the correct combination of movement directions for the elevation and azimuth rotators. There are two traditional approaches to finding this solution. One method involves absolute calculation using the knowledge of the illumination source coordinates, tracker baseplate Euler angles, and the tracker and feedback system optical components to calculate the required tracker movement (e.g. Merlaud et al., 2012; Reichert et al., 2015). The second method deduces the required tracker movement empirically, by taking a small subset of deliberate mispointing measurements to calculate the required tracker angular movements (e.g. Gisi et al., 2011). We used the first approach to generate the basic algorithm, and then only once, when the tracker is first installed, manually perform deliberate mispointing to characterise the optical geometry of the installation in terms of two parameters.

The “translation” parameter accounts for the overall effect of the many reflection and projection translations that occur through the tracker and the optical feedback optics. For example, a single lens might invert an image, and a mirror may perform another translation. The translation parameter is easily determined by intentionally commanding the tracker to shift the image in a certain direction. This can be done using the application's buttons on screen (see Sect. 5.1, which discusses the web server, and also Fig. 6). The parameter string is chosen to encode the required translation so that movements commanded to left, right, up, and down work in the correct relationship to each other. The options for this parameter are “LR”, “UD”, “LRUD”, or simply “N” for none. Once this parameter is set, the “rotation” parameter is chosen to map the resulting image movements to the correct direction on the feedback plane at the diodes. To do this, the observed angular offset is estimated in degrees (e.g. 60) and forms the rotation parameter. The rotation parameter need not be very accurate, within 10∘ is adequate. This is because any small trigonometric errors that accumulate are eventually corrected as if they were from a mechanical source. To reduce a mispointing signal imbalance at the diodes, the algorithm may move a combination of the azimuth and elevation rotators. This is achieved by incrementing or decrementing the error offset variable relating to each rotator. These error offsets effectively become the long-term integrator function, and mainly correspond to the systematic error in levelling and alignment. Error plots of consecutive clear days show similar behaviour, as do plots of similar day length even 6 months apart (Fig. 12), opening up the possibility of using recorded error offset values to identify alignment errors or indeed pre-correct errors during passive mode operation.

Figure 4 identifies the optical components and their pointing vectors. The algorithm was developed to reduce the pointing error by using the mispointing vector from the edge detection diodes. Coordinate matrix transformations T, due to tracker mirror reflections and rotations of the incoming radiation, are required to translate the mispointing vector (Ex,Ey), in the optical feedback plane to the tracker azimuth and elevation axis angular movement reference frame (Eaz,Eel). (Eaz,Eel) = T(Ex,Ey). Unlike Merlaud et al. (2012); the tracker offset Euler angles are not considered in the pointing vector coordinate transformations. With adequate tracker levelling and optical alignment, the tracker baseplate Euler angles are minor and can be compensated for by active tracking. The devised algorithm is mathematically equivalent to that described in detail by Reichert et al. (2015). The readers are directed to this reference for a detailed explanation of the coordinate transformations.

A new solar tracker must be correctly aligned and levelled prior to initial adjustment of the feedback optics. The optical feedback system should be disabled during this process. The solar beam from the tracker must be made vertical (checking by back-reflecting off an oil bath) and the 45∘ mirror (M4 in Fig. 4) under the tracker adjusted to direct the beam horizontally to centre the solar image on the spectrometer input aperture. The tracker and spectrometer optical axes are now co-aligned. The next step is to centre the solar image on the feedback diodes using the adjustment mirror (Fig. 3d). The sky should be clear and cloud-free so that amplifier gains can be adjusted make the four diode values equal. The threshold and hysteresis parameters can now be set as described above (Sect. 3.2). The optical axes of the tracker, spectrometer, and optical feedback are now co-aligned. When the feedback system is enabled it will maintain the solar image centred on the spectrometer entrance aperture regardless of small errors in the solar tracker alignment or levelling.

Software crashes, power failures, mechanical fatigue and exposure to extreme weather are common causes of tracker failure. Choosing a stable and mature OS is a good route to reliability. The OS and computer platform could even be of an embedded nature, removing the need for regular software updates which are an ever-present source of interruption. An uninterruptible AC power supply should be used. The tracker needs protection from weather. This is achieved using good design practices such as waterproofing any electronics. The use of a custom-designed automated cover is recommended. (e.g. Heinle and Chen, 2018).

This leaves us with the main source of tracker failure: the eventual fatigue of wiring and switches that experience daily movement. Our design eliminates this point of failure by having no moving wires and or switches.

The upper rotating stage is comprised of the elevation rotator, mirror, and electronics and needs less than 10 W of power. Traditionally, this has been supplied via a flexible cable or slip rings and carbon bushes – all potential sources of failure. In 2007 we experimented with power transfer using a coaxial transformer consisting of a stationary primary winding with the secondary nested within and able to rotate freely. The primary can then be fixed to the tracker base, and the secondary is fixed to and rotates with the upper elevation stage.

Figure 5 shows this prototype. Early versions used a dual primary with outer (Fig. 5a) and inner (Fig. 5b) coils, with the secondary (Fig. 5c) able to rotate within. This arrangement made the transformer more efficient and capable of transferring more power than was needed, although at the expense of reducing the available optical aperture. The final version of this transformer uses just the outer primary and has an unobstructed aperture of 150 mm. Power transfer efficiency is approximately 20 %. The external diameter of the transformer is 175 mm, designed to fit within the chosen model of azimuth rotator. The primary former (Fig. 5d) also acts as the tracker's base.

The air-cored transformer is inefficient at the low frequencies of normal AC mains (50–60 Hz) but transfers adequate power at frequencies towards the limit of human hearing. A frequency of 15 kHz was chosen as a compromise between efficiency, acoustic noise, potential for radio frequency emission, and physical size of electronic smoothing components. The primary coil is driven with an oscillator feeding a power audio amplifier integrated circuit (IC) type LM3886. Power transfer efficiency is improved by resonating the secondary winding with a suitable series capacitance prior to rectification at the elevation power supply. In 11 years of operation, this method of electrical power transfer has never failed.

The complete tracker was tested for emissions that cause electromagnetic interference (EMI). Between 10 Hz and 75 MHz, the maximum emission seen was -119 dBm Hz-1. No interference was detected from the coaxial transformer.

In addition to greater reliability than other methods of power transfer, the coaxial transformer has the advantage of allowing continuous 360∘ rotation in azimuth. This avoids the need to reverse and re-initialise the rotator daily. In polar regions, measurements can continue uninterrupted through the summer polar day. The elevation rotator and mirror can also rotate freely through 360∘ with no mechanical obstructions. After sunset, the mirror continues to track the solar position (even below the horizon) so that it is ready for sunrise the next day. Except for power-on initialisation, neither rotator needs parking again.

Except for the AC power switch, no mechanical switches are used on the complete tracker. The ever-reliable stepper motors are used for moving the rotators. Park or zero position detection within the rotators is via contactless sensors. No electronic failures have been experienced over 11 years in our four trackers.

Bidirectional wireless communication is needed for the elevation stepper motor controller. We investigated transmission by optical transducers and by modulating the power circuit through the coaxial transformer. Neither method proved easy. Instead we use a generic Bluetooth serial link (e.g. Roving Networks, USA, model RN240F).

Our tracker hardware is OS agnostic. The early version of the application was written in Visual Basic and ran on Windows. However, the short lifespan of Windows versions, along with in-house security rules, soon caused our application software to be obsolete. The application was rewritten in Python 2.7, with the user interface (Fig. 6) in the form of a web server using the Python web framework Tornado (https://www.tornadoweb.org/en/stable/, last access: 2 November 2020). All of our trackers now run on recycled laptops running Ubuntu Linux OS. The tracker can be monitored and controlled from the laptop by browsing to the internal web server. The laptop can also be connected via a second network ethernet card to the spectrometer's PC. Thus, the tracker can be monitored and controlled from that PC too.

To some extent, the tracker, Linux OS, and Python application can be thought of as a stand-alone hardware device with an embedded OS. These should never need upgrading until hardware (the laptop) fails. When this occurs, the OS can be reinstalled or a later version can be used. The tracker laptop and hardware do not have to connect to any other device. Because no internet connection is required for operation, the complete system is largely isolated from the outside world, and thus it should prove very safe from common security threats. Our tracker hardware and software has also run successfully on the popular Raspberry Pi platform. We have not attempted to run the current Python application on a Windows OS. We have not attempted to run our Bruker spectrometers on a Linux OS. At present, our preference is to keep the tracker and the spectrometer on separate PCs.

Figure 7 shows a schematic of the overall tracker and the connections between the major modules. Figure 8 shows the inside the main electronics box within the laboratory. The ethernet to serial convertor (Fig. 8a) sends one channel (elevation) direct to the Bluetooth module (Fig. 8b) and the other channel to the azimuth motor control PCB (Fig. 8e) with the stepper driver daughter PCB (Fig. 8d) plugged in on top. The coaxial amplifier and oscillator, along with associated voltage regulators, are included on one PCB (Fig. 8f). This PCB supplies the regulated direct current (DC) voltages needed for the ethernet-to-serial convertor and motor controller. Analogue signals from the optical feedback system (diode signals) connect to ADC channels on the motor control PCB.

The most likely sources of failure in the electronics bin would be the fan used to help cool the amplifier IC and heat sink (Fig. 8g), and the AC to DC power supplies. In the case of fan failure, risk is mitigated using an extra-large heat sink, which in ambient lab temperatures would suffice on its own. The fan used is a high-quality unit with magnetic levitation bearings (Fig. 8h). Generic DIN-standard AC-to-DC power supply modules are used (Fig. 8c), enabling easy replacement in the future. The total power consumption of the tracker is 55 W.

Figure 9 shows the inside of the elevation electronics box. The coaxial secondary winding feeds a custom rectifier PCB (Fig. 9g). This circuit needs to cope with the voltage extremes and high frequency present in the low-efficiency, high-impedance transformer waveform. A switch mode regulator (type LM2592HV, Texas Instruments, USA) handles the high voltage and keeps power usage to a minimum. The series resonance capacitor is seen at Fig. 9h.

The motor controller PCB (Fig. 9c) is identical to the unit used for the main electronics, good practice for keeping spare parts in common. The Bluetooth module (Fig. 9a) is also identical to the one used in the main electronics. A GPS module (Fig. 9d) and temperature sensor (Fig. 9b) connect to spare inputs on the motor controller PCB. Figure 9 also shows how in this version tracker the elevation mirror adjuster micro-thread screws (Fig. 9e) are placed inside the box. This helps prevent accidental adjustment.

This is the most complex of the modules used in our tracker. The design uses the very popular PIC18F252 microcontroller IC (Microchip Technology, USA). Although our circuit design dates back to 2006, this IC is still readily available (2020). Firmware is written in C using the freely available development environment MPLAB^® X IDE (Microchip, USA). The firmware code runs in a loop, awaiting a command to be received on the serial port. Commands include requests to read the GPS data, temperature sensor, and other analogue and digital voltages on auxiliary pins. Another group of commands deals with the stepper motor control. These commands range from simple step commands to more complex routines such as parking. All commands are acknowledged.

The job of moving the stepper motor is performed by the plug-in daughter PCB – the stepper driver (e.g. Fig. 8f). We made a separate PCB for two reasons. Firstly, this type of driver IC seemed a likely candidate to become obsolete, and secondly this type of IC can be destroyed if abused, for example a short circuit on the cable or unplugging when in use. The IC chosen, type A3979 (Allegro Microsystems, USA), is still readily available as of 2020. Allegro make numerous similar stepper driver ICs and hobby-electronics suppliers use these (and other manufacturers' drivers) in easy-to-use modules for robotics and 3D printers, etc. There is little doubt that suitable stepper motors and drivers are going to be available for decades to come.

We have two methods for monitoring accuracy of our trackers. The first is simply by observation. For this we project a sample of the tracker's vertical beam with a very small (∼ 2 mm2) chip of front-surface mirror (Fig. 3a). Aimed at a distant wall, this acts like a pin-hole camera producing a reasonably well-defined solar image if projected onto a white surface. The circular image can be outlined in ink and any deviations can be assessed in terms of relative movement of the solar diameter. An error of more than 5 % (0.025∘) is very apparent, yet in practice this is never seen. In a similar manner, the focused solar image at the FTIR entrance aperture can be viewed if the instrument source compartment has a transparent cover. Our four FTIRs have these covers.

The second method for assessing accuracy involves post-processing of our routine TCCON FTIR measurements to assess solar-telluric wavelength shift (S-G shift). S-G shift is obtained from analysis of solar absorption features (so-called Fraunhofer lines) in the oxygen column retrieval in the 7882 cm-1 band (Wunch et al., 2011). This analysis uses Doppler wavelength shift of sunlight caused by the relative rotational velocities of the sun and Earth. Absorptions within the solar atmosphere show greatest shift along the equatorial edges of the solar image where the apparent velocity is highest. The main limitation with this method is that sensitivity exists only in the direction normal to the solar polar axis (Reichert et al., 2015). However, the cause of any tracker error is unlikely to fall exactly (and only) along the polar axis for long periods and the solar image also rotates at the spectrometer as the day progresses. This analysis remains very useful. The method does require the use of a spectrometer capable of acquiring suitable high-resolution solar spectra and the user must have the skills to perform the analysis involved. The results are not in real time. Correct analysis relies on the solar image beginning centred on the spectrometer aperture. If this is not done, then the analysis may show a high pointing error even when the tracker is performing well.

We perform S-G shift analysis as part of routine TCCON data processing soon after the day's data are uploaded to our server, and the results are displayed as a web plot for easy viewing.

Figure 10 displays S-G shift on a clear-sky day during an intercomparison between our two Bruker IFS125HR spectrometers (with solar tracker names “Tracker1” and “Tracker2”, based at Lauder, New Zealand). With perfect sky, clean mirrors and both spectrometers well-aligned, this plot shows tracker pointing accuracy (for both trackers) better than 0.02∘. A lone red dot represents the first measurement performed at sunrise, before the Tracker1 feedback was locked. The TCCON pointing accuracy requirement of 0.05∘ is easily met.

This sensitivity of using S-G shift as a pointing diagnostic can be tested by intentionally mispointing the tracker by a small amount. Figure 11 shows the results of shifting the solar image by (approximately) 10 % of the diameter (0.05∘) and 5 % (0.025∘). The orientation of the solar polar axis was not known at the time, so mispointing was performed in two orthogonal axes so that a range of positive and negative spectral shift errors would be captured. Figure 11 shows that the sensitivity of this method is adequate to detect such errors.

It is also useful to see the magnitude of error correction performed by the optical feedback system. The error offset variables represent the corrections the feedback system needs to add (or subtract) from the calculated ephemeris for positioning the azimuth and elevation rotators in order to keep the solar image centred within the four photodiodes. The error offsets change slowly over the day and are logged to file. They provide a good indication of the state of mechanical alignment and/or levelling error. Figure 12 plots these offsets for two different trackers during a spectrometer intercomparison. The test was done again 6 months later. Figure 12a plots these offsets for Tracker1. Some change is seen over the period, although part of the shift between date pairs is caused by the operator performing an offset save within the application. This action adds any instantaneous offsets to the setup offset parameters, and re-zeros the instantaneous offsets. The overall pattern, especially evident in the azimuth (AZ) offset, shows the slow-changing nature of the error sources involved. Figure 12b shows a similar plot for Tracker2, with relatively small offsets that indicate good alignment and/or levelling. Because both trackers indicated a similar (and good) level of accuracy at the same time (e.g. Fig. 10), we can infer that Tracker1 is not well aligned or is no longer mounted level (or the beam was not vertical when the feedback optics were last zeroed). This reinforces how effective our simple optical feedback system can be – even large errors are compensated for by the feedback system. However, it is worth mentioning that if the errors get much larger than the solar diameter (0.5∘) then this feedback system may not automatically achieve lock (because in some situations, the solar image could miss all diodes). Tracker1 is close to this situation and probably needs re-levelling or alignment soon.

A solar tracker long-term accuracy and stability assessment was performed by identifying periods of clear sky (using all-sky camera images) and plotting the analysed S-G shift daily averages from measurements made on these days. Since FTIR measurements are taken automatically, they may still include some observations affected by cloud (or other weather conditions such as haze or wind). However, our current TCCON spectra acquisition and processing procedures have further quality assessment and quality control (QAQC) measures to eliminate the majority of weather affected observations.

S-G shift analysis is also affected by events not related to the tracker. For example, any adjustments (planned or accidental) to the spectrometer perturb the instrument. Because the spectrometer is supported by numerous coil springs, even simply leaning on the instrument or bumping it results in moving the optical axis away from its previous location. The effect of this makes the solar image no longer centred on the entrance aperture. In such cases, the solar image can be manually re-steered back onto the centre of the entrance aperture using the 45∘ mirror underneath the tracker tower. A laboratory logbook records these events.

Figure 13 displays the S-G shift over a 24-month period when Tracker1 was in use. Days with high standard deviation are the result of cloud either affecting a measurement directly or delaying the initial feedback lock of the tracker (and thus are a true pointing error for these few observations). High standard deviation occurring on 1 May 2020 is the result of the intentional mispointing experiments previously discussed. Often logbook entries correlate to a step in S-G shift, especially when the 45∘ mirror underneath the tracker is adjusted. Other large steps in S-G shift, for example in November 2018 and in early January 2019 have no logged cause but are likely the result of human presence in the laboratory resulting in the instruments being bumped. Thus, these steps appear to show the tracker exhibiting poor performance when in fact it is the adjustment of the 45∘ mirror under the tracker or movement of the spectrometer at fault. We have since moved our optical feedback system to be bolted directly to the front of the spectrometer. It appears to compensate for movement of the spectrometer very well but will need to acquire a further time series of measurements to determine the true extent of a system-wide improvement.

We have experienced few interruptions due to failure of our solar trackers. Typically, the trackers run continuously for many months with no operator intervention. The only failure seen appears as a stalled application. To date we have been unable to locate a common cause for this, but suspect it arises from code used in our application or data loss in either the ethernet or Bluetooth links. To put this in context, the spectrometer and PC require restarting much more often than our tracker. The Newport rotators are of excellent quality and are running at such low duty (one rotation per day) that we do not expect them to wear out for a long while.

Development started in 2006, with the first tracker in constant use since 2009. Our most recent unit was built in 2016–2017. The basic design has remained constant, with the only significant change being recoding of the software from Visual Basic to Python. A range of spare parts has been kept: for example, a complete spare electronics bin has been made and is kept at our remote Antarctic laboratory, along with a spare laptop. The electronics bin is completely interchangeable with any other tracker. The motor controller PCB is compatible with azimuth and elevation electronics, and a range of other PCBs are kept as spares. A spare pair of Bluetooth links is also kept on-site at Lauder. We believe our design demonstrates longevity.

Raw feedback diode values are logged approximately each second. These data files are downloaded and used in the spectra quality control process to automatically cull measurements made during periods of cloudiness. In addition, the automated scheduling software used to take FTIR measurements (Geddes et al., 2018) can inspect this file in real time and will only initiate observations when the sky is not flagged as “cloudy”. Using these methods, we have increased our measurement density and quality while at the same time reducing the manual effort required in pre-processing quality control. The web server user interface has proven easy to use. When networked, the tracker is easily controlled from a remote location.

The Python application and electronics can be adapted to drive other types of ALT-AZ trackers: for example, we successfully reused an existing tracker dating from the 1980s. A configuration file allows for alternative resolution rotators to be used.

Our current tracker accuracy is limited by the stability and optical resolution of the feedback system. Although the current feedback optics meet our accuracy requirements, they are far from perfect. Further experimentation using improved mechanical stability and better optics could well be beneficial, especially for when the tracker must be operated in dead-reckoning mode (e.g. using moonlight as a source).

Tracker levelling accuracy could be improved using fine-pitched adjustment screws on the tracker tower baseplate. The use of a permanently attached digital accelerometer may enable easier levelling and ongoing monitoring.

Mirror stepping movements can be perceived in the reference solar image projected to the distant wall, although at times it is no worse than atmospheric turbulence. This movement is the result of the approximate 1 s loop cycle, meaning multiple steps are often needed to track the sun, resulting in a larger (accumulated) single movement than if done several times per second. The present generation of motor controller has rather verbose commands and is quite slow in stepping. The Python application and the firmware of the motor controller have good potential for improvement in this area, especially if greater movement resolution was thought necessary.

There is great potential to add more intelligence to the decision algorithm. Incorporating some Proportional Integral Derivative (PID) control code would make lock occur quicker and remain stable for longer (e.g. Merlaud et al., 2012). Because of the long integration time and “nudge and wait” approach of our feedback method, it takes 1 or 2 min to achieve initial “lock” once clear sun appears. This could also be sped up by saving and reusing offsets from a recent clear day or dynamically changing step size depending on the magnitude of the diode pair difference. The current code is simple, yet few observations are lost with this deficiency because the sky can remain unsuitable for optimum measurements for long periods, and as the sky gradually clears, the tracker is continuously decreasing the pointing error. Once the sun is finally clear, lock is quickly achieved. The algorithm that averages the diode values and uses the threshold and hysteresis parameters could be improved. For example, signal strength history could be used to auto-set the threshold and hysteresis parameters. This would automatically compensate for the reduction in signal level under uniform hazy sky or as mirrors slowly accumulate dust between routine cleaning.

There are clearly potential benefits for switching to a camera-based feedback system, for example it should be easier to perform solar limb measurements (intentionally pointing off solar centre to separate solar and terrestrial absorptions). The tracker has been successfully used for measurements using moonlight, but this was done in passive (non-feedback) mode, requiring careful setup and manual adjustments every hour to account for inherent misalignment and levelling of the tracker. Lunar measurements with optical feedback would be much easier using the extra gain of a camera-based system. If a camera was used, we could retain the philosophy of correcting for a slow-changing error function (“nudging” the error offset variables). An inherent weakness with the edge detection method is the response to a partially obscured solar disc. Under this situation the pointing is affected to some degree but less so than would occur with a quadrant detector. A camera should be able to behave better in this situation.

A camera system, if imaging the spectrometer aperture, provides the important direct coupling of the tracker and spectrometer optical axes that our current system lacks. Alternatively, for critical applications, the current optical feedback system should be moved to attach directly to the spectrometer, effectively tying the optical axes of the tracker and spectrometer together.

Our main code is written in Python 2. A move to Python 3 would be another worthwhile investment.

With the tracker design working well, the next step was to design a matching automated cover (Fig. 14). This needed the ability to rotate continuously in azimuth and to protect the tracker from extreme weather. The cover was completed in 2013 and its use has resulted in a significant increase in our number of measurements. The authors would like to present the design of this cover in a future publication.

It would now be practical to produce much of this solar tracker using modern 3D printing processes. The mechanical structure, including the coaxial transformer components and elevation electronics, should print easily, enabling the design to be easily replicated or scaled to suit, especially for less demanding or cost-sensitive applications. It would be interesting to print semi-adjustable mirror mounts and rely even further on optical feedback to maintain accuracy. The elevation rotator might be a direct-drive stepper motor with a higher micro-stepping value (256 is now possible).