11.2 Solar prediction
Solar prediction using large numbers of high-quality data streams is hardly a novel task, and the performance of existing studies suggest medium-term10 predictions are sufficiently accurate for estimating the coarse behaviour of solar panel outputs. More challenging is the development of prediction mechanisms that operate with minimal data streams. This section is an excerpt from CSIRO’s work , conducted under an Australian Government funded project as part of the Asia-Pacific Partnership on Clean Development and Climate, which addresses this challenging problem, harnessing advanced recurrent neural networks to generate medium-term predictions based only on sampling data available to any standard solar power system. The consequence is a predictive system that requires no additional sensing equipment and, as a consequence, carries reduced cost and complexity for real-world roll-outs.
11.2.1 A very quick introduction to neural networks
Though artificial neural networks have gained much mainstream attention due to claims of simulating human brain activity, their operation is in fact much more straightforward and echoes, at best, only the very simplest components of how brain networks operate. In brief, in an artificial neural network, a collection of input neurons are mapped to a collection of output neurons via some set of intermediary (hidden) neurons. The neural network ‘learns’ by reinforcing the weights of connections between neurons when outputs match user requirements. So, a rain predictor may have an input neuron
that states whether a day is cloudy and an output neuron that says whether the predictor thinks it will rain. Across some set of training examples, the neural network should learn to form a strong connection between the cloudiness of a day and the chance of impending rain. How the network learns such connections ultimately forms the core of most artificial neural networks research. While it is beyond the scope of this report to go into much depth here, note that the majority of successful prediction studies are based around minimising predictive errors across a large number of historical cases and endeavouring to integrate history and context into neural network algorithms.
Importantly for the solar prediction domain, once a network is trained on some set of data (noting that training may continue at any point and with new or incoming data), a neural network is incredibly lightweight. In essence, once input data is supplied or calculated, any prediction is simply one single graph traversal to the output neuron. For most neural networks, this process can be performed on barebones systems with a tremendously fast turnaround that, for the solar prediction task described here, should never exceed more than a few seconds.
11.2.2 Solar prediction with minimal information
In life, it is often the case that the more information we have to make decisions, the better those decisions. And so too with artificial neural networks. The challenge in making predictions for a solar-system with minimal communication and sensor requirements is that there is precious little information from which we can draw conclusions. The goal then is to leverage the data that is absolutely available — in this case, solar power measurements from pre-existing inverter systems.
As expected, forming forward solar predictions based only on the current solar power output is fraught. Even a relatively sophisticated neural network with two hidden layers (see Figure 146) and the effective and contemporary resilient training method fails to produce predictions that are, on average, within 30% of the actual output (using a year’s worth of training data with 15 minute granularity and testing with 60 days of unseen data). Clearly, this is insufficient for most any application of the data and particularly in areas where load shedding and battery charging decisions hinge on at least relatively accurate assessments of future on-site generation.
Taking the lead suggested by contemporary prediction research, this report suggests augmenting the network to capitalise upon history and context. If it is too much to expect a neural network to understand a complex weather system from a single snap-shot, perhaps the performance might improve if that snapshot comes with a story. For example, in addition to the current solar power output, the advanced neural network includes the average output produced over the last hour, the total output produced today and the total output produced yesterday. These simple additions, which carry practically no overhead (both with respect to hardware and software requirements), begin to elucidate trends in the weather pattern that the neural network may be able to capitalise upon. Moreover, the trends exist across multiple time-frames — from the immediate to longer-term.
As a further augmentation, the network also includes a sensible heuristic available to most anyone who is familiar with solar power. Specifically, solar output across a day tends to follow a positive sinusoidal curve which essentially maps the movement of the sun from East to West (see Figure 147). For every time step, the neural network is therefore supplied with a value that indicates where on this sine curve the current measurement was taken. Again, this supplies the neural network with context, providing it with sufficient information to assess whether the solar output is likely to increase or decrease from its current level.
The augmented advanced neural network is illustrated in Figure 148. After resilient training on a year’s worth of solar data, the predictions across 60 days have, on average, improved to being within 18.7% of the true output (with an inter-quartile range between 26.7% and 5.7%). If the network is further augmented to include historical layers that capture decision making data from the previous two time-steps (see Figure 149), the accuracy further improves to 17.7% (with an inter-quartile range of between 25.3% and 5.5%).
This final accuracy level is impressive, but is perhaps best captured through illustration. See, for instance, Figure 150, which demonstrates the neural network’s predictive power across 30 days. Note that the model only produces one poor prediction, with outputs for remaining days successfully capturing the smooth curves associated with clear skies and the more erratic curves familiar to cloudy days. That the neural network performs so well, despite a lack of data sources, is impressive and suggests that supplementing live solar data with static weather forecasts or historical weather data (which could be supplied to the model without excessive hardware costs) may yield results of very high fidelity indeed. Such further augmentations rest as an exciting topic of future work.
Inputs are: the current solar power output; the average solar power output over the last hour; the total solar output across the day so far; the total solar output yesterday; and an indicative point taken from the sine curve in Figure 147.
Inputs include output and decision making memory from the previous two time steps (t-1 and t-2) and the inputs described in Figure 147.
Providing surety and accuracy in predictions of generation behaviour facilitates methods for more intelligent control of resources. In particular, by understanding the likely future state of the solar plant, more sophisticated choices can be made in the areas of battery operation and load shedding. This section has shown that accurate prediction methods can be constructed for renewable generation with a two-hour forecast window.