Leverage reporting

I have spent most of my time at sea on night shift. And one of the biggest things on nights is the midnight marine report (coffee is the second big). Every time data goes in, then it is being posted to all the recipients and then it gets forgotten. New report comes - and that is all that matters. However, working in the Gulf of Mexico we have collected some good set of data - at that time around two years non-stop at one location (Mississippi Canyon). And when a time had come for a new landing of BOP, my boss - Master/OIM Paul Morris - said: "Roman, maybe you can come up with some sort of wind chart?" And so it started.

A few days non-stop collating reports into one excel sheet, cleaning data, figuring out how to plot it correctly - and the result has been sent out for judgement. "Very good!" he said. That was probably my first project with data and I liked it. For the change in routine and for the sense of accomplishment. After that I have learned how to do it faster, produced a few more for presentation to HESS and for decision-making with BOP landing.

As you live - you learn, and more tools are getting into our toolboxes. Same happens with data, in my case - thanks to Python - I have never lost interest. On the contrary - it became more fun. Serialisation of data has a very big potential. Frequency of sampling outlines the area of analysis. Our midnigt reports are man-made, data is sampled every hour at best, and to get 1000 readings - statistically significant - you need more than a month of records. This can be used to tell the story of one well, or even the field.

More frequent sampling is beyond human ability and can be performed by computers, more on this in the next article.

So, after changing a few units in the company, I decided to have a look at some past data. It covers a period from July to September inclusive. This time, instead of focusing on winds only, the whole hourly log has been collated, exported to CSV format and fed to Python. Resulting data frame had 25 columns, and described like so:

Doing just that we can have an idea of what sort of conditions the unit (drillship) experienced while on the field. This is close to what could possibly be done in excel, but the advantage is scalabily - any CSV of similar format can give this data in a matter of a minute. Moving further, we can plot the set and have a glance at what areas might be of a particular interest.

Looking at this representation, sharing one X-axis (which is date and time), one can distinguish similar behavior in several time series. To check our intuition out, we can compute this and plot. In this case, a value showing how much two series related is correlation. We will calculate correlation matrix for our frame and then show it in a nice form.

This is getting more interesting, as we start to see the representation of our intuition. The chart shows that - indeed, some series are correlated. Brick-red cells show positive correlation around 0.25, Marine-blue negative 0.7 (which is quite strong).

Let's have a look what it is. Sats (number of satellites currently used bu the DGPS) negatively correlated with HDOP (horizontal dilution of precision). That means - the more satellites we have, the better is the quality of positioning (small HDOP is good).

Lower left is bar - barometric pressure. It is in its turn negatively correlated to a lot of series: wind speed, pitch, roll and heave, waves and swell. That means - if pressure drops, weather gets worse in general. Wind speed increases, causing higher seas, which cause our ship to move with higher amplitudes. DP Force - used to keep us in position, reasonably positive to the increase in external factors. To see what is the distribution of two related series we can even plot a kernel density map for them:

We can see though that some dependancies are stronger than others. It may be explained by a certain time lag between the occurences (a sudden wind gust does not necessarily mean high seas straight away). But that relation in general gives us one interesting ability - knowing the history (having statistically significant number of observations) we can predict, to a certain degree of precision, how our vessel behaves on theoretical well location, given the expected weather conditions. To do that, we will need to get better data (for increased precision), feed it to neural network, train it and save the model. Then feed new data (without target variables - pitch/roll/heave/force) and get the result.

More on modelling - coming soon. Thanks for reading.