Natural Gas Storage Optimization : An Entry Point for AI?

  • Shawn Smith
  • Published: 03 July 2024

The adoption of Artificial Intelligence (AI) in the energy sector is in its infancy, but many innovative energy companies are exploring ways of using the technology to improve commercial operations. In this article, we explain why natural gas storage optimization is one of the energy sector's most promising AI use cases. 

Natural gas storage optimization involves maximizing profits and operational efficiency by timing natural gas injection and withdrawal schedules. However, these priorities depend on an organization’s business perspective.

An industrial-scale consumer, such as a power generator or utility, that is holding storage capacity will focus first on the surety of supply and then on minimizing costs. On the other hand, a trading company will commonly leverage storage positions to improve trade margins with strategies that consider current and future market conditions.
Setting aside the differences between these participants, gas storage optimization programs focus primarily on current and forecasted prices. Are prices too high to inject? Are prices too low to hedge? From a trading perspective, the goal is to minimize costs when purchasing gas or by hedging during injection season and then maximize value during withdrawals. 

Optimization also implies active management of both hedges and physical movements of gas into and out of storage, taking advantage of market changes to capture near-term (spot or intra-month) value while ensuring the asset is properly positioned to execute longer-term seasonal strategies.

Optimization today

Currently, companies seek to optimize storage positions using one of two methods. 

The first is essentially an intellectual exercise where traders responsible for the storage book review current and future market conditions and physical positions stored in the organization's ETRM systems and dedicated spreadsheets. Using analysis and intuition, they manage that position, including potential financial hedges, to extract the highest value. However, reliance on one or more experienced traders does create risk. 

Without a programmatic approach that facilitates backtesting and benchmarking, it is difficult to know if the trader truly optimized the storage position. Put another way, we can never say with certainty how much money they might have ‘left on the table’. 

The second method is to use analytical models. These may be internally developed in systems such as MATLAB or provided by a specialist risk solution vendor such as Lacima or cQuant. The most sophisticated models use Monte Carlo simulations to determine optimal injection/withdrawal schedules based on various inputs, including seasonal forward prices, cost of fixed and variable storage, cost of transportation, and capacity constraints.  

Unfortunately, incorporating the best commercially produced systems can be difficult and expensive. As for sophisticated, internally developed models, non-data scientists may struggle to understand them and their limitations and lack the ability to maintain them in the long term.   

In either case, even with a good model, the outputs are only as good as the data it consumes, and these data challenges are difficult to overcome. Issues include latency in collecting relevant data and human or interface input errors. 

Additionally, published market prices and indices may lag behind sudden or near-term developments such as unscheduled pipeline maintenance, accidents, severe weather events, or geopolitical developments such as wars or unexpected tariffs. All these factors can significantly impact the accuracy of model results. 

Other data considerations that may be less obvious but also important include the creditworthiness of counterparties when opportunities to sell out of storage or buy supply present themselves. After all, what is the point of running a model for a particular market opportunity if the sources of supply or markets cannot clear the credit hurdle?

Is this an opportunity for AI?

AI has shown it can consume vast amounts of data and return answers to complex problems in several industries, including fraud detection in financial institutions, piloting autonomous vehicles, image and facial recognition, and even medical diagnosis and prescription.  

Given these capabilities, can AI become a new natural gas storage optimization tool? Possibly, but there are still plenty of factors to consider. 

Given that forward prices are perhaps the most critical factor in an optimization model, using AI to predict prices is potentially the most valuable opportunity. However, it still remains the most challenging task to perform accurately, especially given the current impact and future unknowns of the energy transition away from hydrocarbons. 

However, AI could enhance the market awareness of traders by consuming not only the common operational considerations such as MDQ for injection/withdrawals, costs, pipeline capacities, and schedules but also less structured data and information that influences the shape of forward price curves.

Accurately forecasting future prices is challenging, as we saw above when highlighting the role of unforeseen events, including weather patterns, geopolitical events, and economic indicators.  

Moreover, if one were to simply look for the financially optimal solution for that day's gas flow without considering the potential variables cited in this article, the indicated ‘financially optimal’ solution may increase operational risk in that gas flows may be cut or curtailed under some types of transportation agreements. 

Nonetheless, given the ability to rapidly update in changing market conditions and provide near-continuous updates, an AI model could provide some incremental benefit over existing models or methods. And this is not only for price forecasting – AI can even suggest a trading strategy that could be used for gas storage optimization or trading in general. 

AI would learn from historical data on the market environment, encompassing not just prices but also factors like pipeline utilization, market events, and storage technical conditions, along with historical traders' decisions, and provide traders with direct decision support. This kind of AI-based trading is already extensively used in financial trading and can be adopted for energy trading in the near future.

If AI is successful at improving the accuracy of forward price predictions or in creating meaningful trading strategies, natural gas storage is only one activity that would benefit. A model that improves decision-making and the accuracy of future price forecasts would be of great value throughout the trading floor, from origination to C&I retail and beyond. 

AI may also allow us to consider solutions that may not be ‘optimal’ but serve some longer-term interest or goal. These might include ESG considerations or counterparty factors (such as creditworthiness or reputation). 

Complexities and opportunities of AI in natural gas 

There are other issues to consider. Accurately predicting forward prices for any location could leverage AI models to consume a vast array of information and develop an understanding of the magnitude of influence of each of the various inputs without relying strictly on history, as current market developments are often without precedent.  

Further, bias in the collected data is a significant concern. Many inputs are subject to the influence of an unstable political climate and messaging, editorialized press coverage, and social media. So, understanding where these AI models are harvesting data, how they are weighting that data, and ensuring the elimination of bias wherever possible are vital to ensuring accuracy. Unfortunately, in the near term, reliance on these models for price forecasting will come with risks until they acquire a track record that can only come with time. 

Also, it's worth considering that there are almost certainly other actors, such as global-scale commodity hedge funds, which are investing heavily in the same capability. Should these large-scale firms actively trade based on the results of their proprietary models, their actions might induce a price feedback loop that influences forward prices in unpredictable ways while creating and amplifying volatility that feeds into the next iteration of the AI forecast. 

Given these considerations, until AI-based price forecasting is a proven and reliable capability, it might be best to consign this use case to the IT lab, where it can be monitored, assessed, and validated. 

In the meantime, energy organizations should look for other opportunities where AI can improve financial and operational performance in natural gas-centric businesses, be it storage optimization, gas scheduling, or upstream and downstream operations.  

Here, the ability to leverage the technology will largely depend on the scope and scale of the variables involved in the particular use case. Those that involve the most significant number of ‘market-influenced’ variables, as we have seen with forward prices, will be more challenging to tackle in the near term. 

Given that difficulty, an early deployment of AI could create a model that considers historical disruptions to the natural gas markets (such as the impact of severe weather events, price crashes, regional imbalances, and perhaps even cyber events). This would essentially provide an ‘early warning system’ of potential catastrophic market events, allowing the user to flatten their positions in time to limit significant financial losses.  

Where does AI go from here?

While the use of AI in energy trading markets is still evolving, the potential for useful adaptation is evident. The decision-making process for trading and hedging involves numerous complexities that may not be able to be fully automated by AI at this time. However, there is a strong argument for using AI to improve traditional trading opportunities and optimization. Additionally, introducing GenAI may be helpful in the valuation of energy assets (including gas storage), but this is a subject for a separate analysis.